目标：统计文章出现频率最高的单词，输出频率最高的前十个单词。
以下是分析的具体内容：
（1）首先是cpu使用百分率：总执行时间为29.3秒，截图如下
（2）各函数占用时间如下图：
（3）程序运行截图：
所需测试英文文章：
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
Camping along the Big Long River
Abstract: Geographic profile is an important way to study serial crime. How to generate geographic profile is sophisticated. The paper proposes a method to generate a geographic profile. The method consists of an analog model and a simulation algorithm based on Bayesian formula. The analog model requires the location of serial crimes to meet the "ring hypothesis". we estimate the offenders’ residences by the physics method to find the minimum distance and find the minimum point of the sum function of the distance between the offenders’ residences and the crime locations.
The main idea of this method is as what follows. Making the point (quasi-residence) derived by the analog model controls the direction of simulation, searching the locations from outside layers to the inside ones, taking the final search sequence of quasi- residence points, solving the minimum point of the function which is the sum of the distance between the final sequence of quasi- residence point searched and the crime sites. The minimum point is the residence of the serial offender.
By BP neural network we predict the next possible crime point. Finally, we test our approach by serial murder cases of Peter Sutcliffe. We derive the result that the distance between the predicting residences point and the actual residence point is only 9.892kilometer, which shows that our approach is feasible.
Key words: ring hypothesis; geographic profile; BP neural networks; Peter Sutcliffe simulation
Contents
Executive summary..…………………………………………………………………3
1. Introduction ……………………………………………………………………...5
2. The factors that affect the serial crimes ………………………………………..5
2.1 The distribution of the population density ………………………………5
2.2 Topographical features …………………………………………………...6
2.3 The nature of the case itself …………………………………...................6
2.4 Factors of the offenders themselves ……………………………………6
3. Analysis of Mike O’Leary’s Method ……………………………….....................6
4. Analog Model …………………………………………………………...................8
5. Simulation Algorithm Model ……………………………………………………..9
5.1 Terms, Definitions and Symbols …………………………………............9
5.2 Assumptions ………………………………………………………………..9
5.3 The theory of Model ……………………………………………………...10
5.4 Methodology ………………………………………………………………11
5.5 How to determine the variable d ……………………………………….13
5.6 Model test …………………………………………………........................ 15
5.7 Analysis of the Result ……………………………………………………19
5.7.1 Sensitivity Analysis ………………………………………………...19
5.7.2Model improvement ………………………………………………...20
6. The combination of the above model …………………………………………...21
6.1model foundation…………………………………………………………..21
6.2 Single hidden layer BP neural network model………………………..22
6.2.1 Methodology………………………………………………………...22
6.2.2 Determination the layer of the network …………………………23
6.3 Model Evaluation………………………………………………………….25
7. Conclusions and future work……………………………………………………25
8.References ……………………………………………………………………….. 27
9.Appendix…………………………………………………………………………..28
Executive Summary
The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, The analysis of the serial crimes’ Geographic Profiling is a relatively mature study direction in the field of crime study. Because we can predict the time、 location and the base point (generally considered to be the living or working place of criminals)that offenders commit the crime next time according to the psychological characteristics of offenders and local factors such as geographic information. In this way, it will help the police to crack the criminal cases. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. But in actual operation, these results are too complicated and difficult to operate. The purpose of this paper is to propose a simple, practical, easy to operate, and high reliability of the prediction programs.
In our models, we have used the physical principles, mathematical calculations, computer simulation and other advanced technology to solve this problem. In some way, we should try to learn to think like offenders. we assume that the crime states is a place where has the maximize chances to commit crimes successfully, at the same time it has the minimize dangerous. On one hand, this increases the reliability of our projections. On the other hand, due to the density of population distribution and topographic factors, the serial crimes should be divided into four kinds of circumstances. As the flowing words:
? If the offender's modus site distributes in a populous and little affected by topographic factors, we have proved that the base of offenders is largely the geometric center of gravity formed by all his sites. This result is a good guidance for the populous large cities searching for serial criminals.
? For the evenly density of population distribution and more influenced by the topography areas, because criminals need to consider travel problems, then we have reason to believe that the base location of offender is the point that the minimum summary of Euclidean distance between the base point and the crime states. For this problem, we have two programs, the second one is a numerical calculation method, the first one is the analog simulation of a physical process to seek this point. We propose to adopt the first option.
? For the not evenly density of population distribution and not influenced by the topography areas, we recommend to use the criminal behavior Model ,developed by Towson University Mike O'Leary, Professor of Mathematical.
? For the not evenly density of population distribution and more influenced by the topography areas, we can use the prior data to train a BP neural network to predict the next location of one offender committing the crime.
So that we using different strategy to make prediction for each case, rather than a general single approach. So it is highly adaptable and promotional. Although this issue has been solved relative successful, a new problem comes out again, that is if the offender understand the above theory, then he will no longer or less commit the crime in such geographical contours and they choose other points as their targets, this time we cannot use these programs in direct.
I. Introduction
In U.S. and Europe, Peter Sutcliffe's serial murder cases cause the study such crimes. In recent years, it has become a relatively new field of study that uses the Geographic Profiling to frame the investigating scope of the serial crime. It is considered as a more effective method in the detection of major, difficult series. Since the early 1990s Kin Rossom officer began to investigate the Geographic Profiling, Geographic Profiling has become a practically useful method in criminal investigating and judicial practicing. There is extensive literature on defining the Geographic Profiling, and recent debates have emerged on how best to combine at least two methods so that a more accurate scope can be found .Combination of methods requires at least two different kind of Geographic Profiling model and a suitable method to make an accurate scope.
So our goal is clear greatly:
? Creating at least two kinds of Geographic Profiling models
? According to the given data, operate the Geographic Profiling model.
? Devise a method to combine the Geographic Profiling.
? write a report to the person in charge of law enforcement agencies
Our approach is:
? Search the literature on existing evaluation methods
? Develop a comprehensive method that asks only for existing
evidences or evidences easy to measure and collect.
? Collect experimental data that can be used in our method.
? Compare current methods and determine their characteristics
? Do a sensitivity analysis of variations of our models.
? Search the literature on existing evaluation methods and find their shortcomings.
? Do further discussion based on our work.
II. The factors that affect the serial crimes
The definition of Serial killer is committing the crime more than three times within a month and the modus operandi is generally the same or similar. So it is very necessary to posit their living sets. Research in this area has taken a broad approach by including in its analyses operational, perceptual, behavioral, physical, social, psychological, legal, cultural, and geographic settings. （【5】D. Kim Rossmo PLACE, SPACE, AND POLICE）. And these characteristics are more likely to become parameters in a mathematical model.
2.1 The distribution of the population density
Serial offenders generally choose populated place as their base stations.
? Selecting densely populated areas can reduce the distance from his base to the crime point ([13] Di Lianhai) .It is also propitious to the criminals escaping.
? There will be more “prey” in the densely populated areas.
? It is conducive to the offenders covering in densely populated areas.
2.2 Topographical features
Serial criminals will not set their base around the lakes, rivers, hillsides, etc.
2.3 The nature of the case itself
Different types of serial criminal sets have a clear geographical distinction .For example, bar robbery or house robbery has a fixed location.
2.4 Factors of the offenders themselves
For the same form of serial criminal, different cases of criminals will have a different geographical distribution
The difference of the economic power and tools for committing crimes would affect the choice of the crime locations.
III. Analysis of Mike O’Leary’s Method
Mike O’Leary considers that all the known crime sets are mutually independent and obey normal distribution. The assumption is under the conditions that all offenders have the same attack distance. According to Bayesian theorem, he establishes a probability density function under the conditions that all the crime sites have known.
He uses the flowing assumption:
? the offender chooses distances according to a Rayleigh distribution, but that different offender / offense combinations have different average distances.
? Not all locations are equally likely to be the sites of an offender’s home base.
? Density of potential offender anchor points is proportional to the local population density.
? The serial criminal offenders have the unique base location, either residence or a working point;
The mainly formula of Mathematical Model:
(1)
The mainly formula of Criminal behavior Model：
(2)
(3)
(4)
(5)
This model includes all the factors that affect the serial crimes.
Strengths:
? The analysis is very comprehensive, and almost all the factors relating with the serial criminal have been included in the model.
? If we give the model sufficient crime sites, it can predict the location of a base point with high accuracy;
? This model can reflect the probability of the base-point falling into one region in numbers;
? It can be combined with GIS platform to develop secondary.
Weaknesses:
? This model is not suit for the newly emerging comic crime type. If we use very little sample points in the model, the accuracy is not high.
? The software will run a very long time to get the best spots.
IV. Analog Model
By Carter's "ring hypothesis": we marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle. (David Canter, Toby Coffey and Malcolm Huntley Predicting Serial Killers' Home Base Using a Decision Support System)
Lemma: first we drill N small holes in a wooden to make an irregular polygon matching along with the crime point. Then take N same weights with the same length of rope through the small hole marked with a summary post. When all the weights are stationary, the location of the node is the point that the summary of this N-point distance is the least.
Figure 1: The physical model in the prior lemma
Proof: When all weight stationary, the location of the node is the point that the summary of this N-point distance is the least.
We have known that when a system is in static state, its total energy reaches the minimum. In this system, there is only gravitational potential energy. So when it is in the state of stationary, the summation of the gravitational energy is the least. So the summation length of the rope under the wood block is the largest. Therefore, when the total length of the rope is a fixed value, the rest is the least. So the location of the node is the point that the summary of this N-point distance is the least.
So we can use the analogy method to find a base of criminals:
? Plot out the N crime sites on the map to make N-gon;
? Look for a board. Then review the board with the map. Finally drill N holes on the site of the N-vertex of the N-gon;
? Found the location that the system in the stationary state;
? Find the latitude and longitude of the node location. And this latitude and longitude is the general site of base.
The use of analogy can be very easy to find the general location of a base point. And then the police can search the criminals around this point.
V. Simulation Algorithm Model
5.1 Terms, Definitions and Symbols
? is the i crime location, expressed by longitude and latitude
? is the ith crime location, expressed by Gauss plane coordinate system
? is the location of predicting persistent base, expressed by Gauss plane coordinate system
? is the angle generated by the computer simulation procedures,
? is the i distance between the crime location and the base
5.2 Assumptions
? Assumes that all serial offenders have the so-called persistent base, in our paper we made the offenders’ residences as the base point.
? We support "ring hypothesis", which marks all the crime sites on the map to find the maximum distance between all crime locations, with the largest distance as the diameter to draw a circle which includes the all crime locations. We assume that the offender living in the circle .Furthermore more, they usually live near the center of the circle.
? In order to simplify the model, this paper only considers the individual committing the crime, does not study a chain gang offenses.
? After the offender committing a crime in the prior incident place, he first returns to his persistent base, and then he goes to the next incident place.
? Among all the occurred crimes, there are reasonable evidences proving that a series of occurred crimes was committed by one person.
5.3 The Theory of Model
Based on the assumption of Mike O'Leary model, using the Bayesian theorem as the theoretical basis of the premise, we establish a computer simulation model based on Bayesian models. First, we use Monte Carlo method generating pseudo-random numbers in order to approach achieve the base point step by step. Using our approach to produce a series of random numbers does not occupy memory after completing the program, the system configuration requirements low. What’s more, for new types of crime, we can predict the base point in high accuracy. This is the inadequacies to the Mike O’Leary model.
According to the all crime locations and persistent base confessed by the serial offender arrested by the police, we can found that the distance between crime locations and persistent base keeps to the following assumptions if we in the light of the serial offenders’ criminal psychology:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals are for fear of being recognized or attracting the police's attention, therefore, generally ,they do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
From the mathematical point of view, the event that criminals select a series of crime bases commit can be seen as independent. Through Mike O'Leary’s Models of Offender Behavior([17] Mike O’Leary Department of MathematicsTowson University 2009) we know that the distribution of the crime base subjects to the normal distribution.
Combining with the assumption at the beginning of this article, we believe that the whole process of serial criminals committing crimes is a radiation network from the persistent base to crime points. As the following figure:
Figure 2, the network between the crime sites and the base point
5.4 Methodology
We get the above conclusion on the basis of analyzing the crime locations and persistent base confessed by the serial offender arrested by the police. Well, to the reverse, if we have known a certain serial number of their crime targets, we need to find the general scope of the possible base of the perpetrators. Through the Mike O’Leary Models of Offender Behavior, we know under the circumstance that we can only find a series of crime points, if the persistent base is exist, the crime points occurring in the X Division is random, then the distance between the crime locations to its base is also random .So we select the distance and angle to develop a model.
Here we have to show why these two indicators can be used to find the location of the crime "center of mass":
Regarding to the selection of indicator : Through the Mike O’Leary’s Models of Offender Behavior ,we know that under the condition that the base is existence and only one ,the location of the base belong to the normal distribution .
As
So
Analogous to the crime base subjecting to normal, under the circumstance that the unique existence of the base, the probability density function is : .
subject to normal distribution, under the circumstance that the unique existence of the base ,the probability density function is : (6)
Similarly, there are:
(7)
is a fixed value , referring to the average distance between all the crime point . In this case, we will use the maximum likelihood estimation to calculate its value.
The selection of : Based on the analysis of the whole criminal process of serial offenders, we have concluded that from the base to the crime point is a network of radiation patterns. Since the crime location subject to the normal, and the occurrence of the crime location and is at the same time, so the occurrence of also subject to normal.
Determination of anchor points:
Firstly, supposing the positioning of point (center) coordinates we are looking for. It is .
Then, we will establish the objective function:
. (9)
Making the smallest point of the function S is that we are looking for anchor points:
(10)
This is the algorithm flow chart which we designed:
Figure 3, the flow chart of Simulation Algorithm Model
5.5 How to Determine the variable d
Use Monte Carlo method to generate a series of distance from the crime location.
During the process to determine the variable d, the "ring hypothesis" is still valid, because the distance from the base to the crime point subjects to the flowing assumptions:
? Under the psychology of "rabbit does not eat the grass around the nest," the criminals do not choose to commit crimes near the base;
? The distance from the base point to the crime location is not too far, because the further the distance is, the lower degree of familiarity the criminals feel. What’s more, it will bring a series of problems such as the unfavorable escape. This is also consistent with Rossmo distance decay function argument.
So that we assume that near the base of criminals, there is with a buffer zone. This region has the following characteristics:
? The buffer zone largely on the form of rings including the base ;
? In serial cases ,the crime points are mainly (more than 90%) located in the buffer zone;
? The crime points felling in the region subject to the normal distribution.
Figure 4, the geographical distribution of the general modus
In the above figure, is the minimum distance between any two known crime location.
This formula is the maximum distance between any two known point.
Our approach is:
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
(11)
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
(12)
is in correspondence with the sample crime point , Here we have narrowed the scope of persistent base included in the scope of , successful narrowed the search range of the center of gravity locations.
? We already obtained a false base distribution in the front step. We must remove the singular point which in the actual problem according to the simulation effect chart some accident factors create. Then we will establish accurate reference point Gauss who retains the place coordinates vector.
5.6 Model test
We use the title of the serial murderer appears in Peter as our object of study, using just the method to determine the location of his points. First looks for on Google earth his committing a crime place such as:
Figure 7,mark the modus of Peter Butcher by Google earth
. We search the information of the Peter series case murder and reorganize the data following three forms:（http://translate.google.cn/translate?hl=zh-CN&sl=en&u=http://members.fortunecity.com/hiper22/sutcliffe_cf.htm&ei=yyx-S4eHFYro7AOM75zlCw&sa=X&oi=translate&ct=result&resnum=10&ved=0CCkQ7gEwCQ&prev=/search%3Fq%3DPeter%2BSutcliffe%26hl%3Dzh-CN%26newwindow%3D1%26s）,here we suppose that the information introduced by this website is objective and real, and may be used to test our model.
1. Data ordering:
This is the form on the data of series murder case such as（Table 1 see in Appendix）:
Use Google earth to obtain the latitude and longitude of earth each guilty place such as:
Table 3: the comparison table between Geodetic and Gauss coordinates of Peter Butcher’s previous modus
name central meridianLatitude B Longitude L
Keighley 183 55.515542 181.543468
Halifax 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Manchester 18353.285947 182.144502
Leeds 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Huddersfield 18353.384099 181.470232
Manchester 18353.384099 181.470232
Halifax 183 53.432944 181.514006
Bradford 183 53.473811 181.450879
Farsley 183 53.483567 181.402361
Leeds 183 53.432944 181.514006
Huddersfield 18353.384099 181.470232
Leeds 183 53.432944 181.514006
Using the execl to transform the latitude and longitude coordinate into the Gauss place coordinates:
Table 4：Gauss place coordinates
X Coordinate Y CoordinateX Coordinate Y Coordinate
6194222 -68266.95947164 -80418.6
5955997 -75173.85947164 -80418.6
5955997 -75173.85955997 -75173.8
5955997 -75173.85963805 -82212.5
5955997 -75173.85965680 -87399.3
5955997 -75173.85955997 -75173.8
5963805 -82212.55947164 -80418.6
5955997 -75173.85955997 -75173.8
5963805 -82212.55918593 -102055
5928763 -50065.65958738 -79053.4
5955997 -75173.85963805 -82212.5
Computer simulation:
Put the obtained coordinate data in the procedure to let the computer simulate and obtain an anticipated base distributed graph to be as follows:
Figure 8, Distribution of the possible residence
The red spot expresses the Peter series case of murder's locating point, the blue color dot expresses his committing a crime place, and the asterisk expresses the accurate locating point simulated by computer simulation. After digging several singular points that don’t conform to the request, you can see that this model can well reduce series committing a crime happening place central point scope to seek for. The graph can well explain the reliability of our model.
3rd, determinate locating point coordinate: from simulated accurate locating point vector: ,
We can get the coordinate of the accurate simulation series points and make these coordinates into form such as:
Table 5: series coordinates of Quasi-analog-point
X Coordinate 5.93465.9120 5.92485.9414 5.9234
Y Coordinate -0.0755-0.0930 -0.0761-0.1130 -0.1050
X Coordinate 5.96185.9168 5.95015.9131 5.9481
Y Coordinate -0.1071-0.0909 -0.0921-0.0536 -0.0753
X Coordinate 5.95075.9391 5.94725.9378 5.9619
Y Coordinate -0.0837-0.0911 -0.0881-0.0912 -0.0877
X Coordinate 5.95115.9351 5.94485.9383
Y Coordinate -0.1280-0.1125 -0.0963-0.0826
Takes the objective function the minimum value:
Then get the result of the final locating point is:
（13）
5.7 Analysis of the Result
From the information about Peter we know that Peter lives in a place north of Bradford about 6 kilometers away. Comparing this distance with the earth radius,6 kilometer is relatively small. Therefore, we think that Bradford is the actual location point of Peter.
The distance between the predicting location and the actual location is:
(14)
The error distance, 9.892 kilometer, is relatively small relative to a residential area. Greatly narrow the search range.
5.7.1 Sensitivity Analysis
Observe the change degree of the base point location when give variable size of .
We choose three larger date (4.75, 5.10, 5.34 ) and three smaller ones (3.50, 3.85, 4.25), enter the 6 data into the Simulation software and get the flowing result in table 5:
Table 5: Simulation results:
3.50 3.85 4.25 4.75 5.10 5.34 4.5681
1.12 1.28 1.22 1.15 1.13 1.08 0.99
1.1412
1.2976
1.2357
1.1596
1.1376
1.0975
1
From the table we can see:
? Whether a larger or a smaller , will make the error distance between the predicting base point and the actual base point become larger. It proves that the accuracy of our computer simulation model.
? The change of is in equivalent of another kind of serial offender’s average attack distance. We can conclude that different criminals have different attack distance in the same kind of serial crimes.
Observe the change in the distribution of pattern after the changing of :
Run the MATLAB program we can obtain the following distributed photo:
Figure 9, Distribution of the possible residence
From the last two diagrams, you can see the distribution of pattern gather to the center. And during the process of simulation, we find that the actual crime point has been included in the distribution of base points. This illustrates that the adaptability of our model is very good.
Comprehensive the two previous factors, you can find that our model has practical operability.
5.7.2 Model Improvement
The computer simulation model is based on Bayesian formula. We conclude that the distance between the base point and the crime sites obeys the probability density function:
on the basis of the probability density function
which propounding by Mike O’Leary.
This is a purely mathematical process, and does not relate to too much geographical information about the crime sites. Although this model can predict the location of the base in high accuracy, it considers too less factors. So we can improve the accuracy. Based on this, we have established the computer simulation model basing on Mike O’Leary’s criminal behavior Model:
（15）
（16）
? Using Monte Carlo probability density function to produce a series of pseudo-random number, getting a random distance from the “center of mass” to the crime point. The formula is :
? Through the computer simulation, obtain a series of pseudo-base-point vectors of "distance" and "point of view" indexes:
（17）
（18）
So that we will be able to narrow the region of the base point, while according to the information of the geography we can get the probability that the criminal falls into the region.
VI the combination of the above model
6.1 model foundation
Use the model 1 and model 2 to generate the final geographic contours:
In practice, we use the model 1 to generate a possible base point
? Use the model 2 simulate to the possible point layer by layer. Each layer the distribution of the patter correspond to a value
, this value is the probability that the point is the base site.
? The probability value indicated by the color of the point will be able to generate a geographic profile, using to predict the possible suspects hiding places, in order to give the information to the police.
This is the effect imagine which we obtained after carried out using MATLAB simulation with the assuming data:
4851
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
