每天一篇好论文

2023/01/29

Learning to Handle Parameter Perturbations in Combinatorial Optimization: an Application to Facility Location

• doi:10.1016/j.ejtl.2020.100023
• priori
• reference instance
• We have shown that the existing or simulated data of a recurring optimization problem can be exploited to gain insight into the optimi- zation process. We have used these data to fit a binary classifier and a regressor. We predict whether a perturbed instance, derived from a reference one, will share all or a part of its optimal solution with the solution of the reference instance. This piece of information, translated into an additional constraint, is given to a solver and allows to dive faster, on average, towards a good solution. Moreover, we empirically illus- trated that this additional constraint preserves the optimal solution and thus prevents negative transfer.

2022/03/13

A beam search approach to the container loading problem

• doi: 10.1016/j.cor.2013.09.003

名词列举

• CLP single container loading problem (CLP)

• strongly heterogeneous (many types of boxes) and weakly heterogeneous (few types of boxes)

• constructive, divide-and-conquer and local-search.

1. Constructive approaches generate solutions by loading boxes into the container until no more boxes can be loaded.
2. Divide-and- conquer methods divide the container into subcontainers. Then, they recursively solve the resultant smaller problems before recombining them into a complete solution.
3. Local search methods start with a complete loading plan, then apply neighborhood operators to generate new loading plans.

block-building-based

• simple blocks

• general blocks
  (K1) the representation of the free space in the container;
  (K2) the mechanism to generate blocks;
  (K3) the heuristic used to rank free spaces;
  (K4) the heuristic used to rank boxes;
  (K5) the heuristic used to decide how a selected block is placed in the selected free space
  (K6) the overall search strategy.

• Beam search 束搜索can be viewed as an adaptation of the branch-and-bound search that expands only a subset of the most promising nodes at each level of the search graph

• State : BSG-CLP works by exploring the search space to find a path from some initial state to some terminal state.

• State : a list R of overlapping cuboids which represents the residual space in the container, a list of the remaining boxes C, and a list of blocks B.

• terminal state (or complete solution) corresponds to a loading plan in which no more boxes can be added. In other words, any box in C fits a cuboid in R.

search space can be represented by a directed acyclic graph (DAG), where the vertices are the states, and the edges are the transitions.

• 放入一个block后,需要更新的东西,1. remaining boxes C 2. B remove unfeasible block 3 residual space 剩余空间 R
• double search effort

2021/11/22

New approaches to nurse rostering benchmark instances

• doi：10.1016/j.ejor.2014.01.039

名词列举

• Rostering problems
• string/pattern matching
  |Grouping|Matching one of a group of shifts at a point in the
  sequence|
  |Negation|Matching anything but a specific shift or group of
  shifts at a point in the sequence|
  | Alternation | Matching multiple patterns |
  |Quantifiers|The pattern(s) must appear a minimum or maximum number of times|
  |Restricting the search text to a specific region of the work
  schedule|
  |Only matching a pattern if it starts on a particular day in the
  work schedule|

结论集合

1. 在RP问题中，常见的算法包括Metaheuristic/Constraint Programming/Mathematical Programming/AI;不同的算法在不同的算例上表现不同，MP在一些算例很快就能求到最优解；
2. NRP Benchmark Instances

One of the most significant challenges in addressing a large diverse collection of instances is developing a model which can be used for all the instances with their varying types of constraints and objectives.(排班模型中的各种复杂约束，通常很难建模表示)

About 5% of the computation time is spent re-solving the restricted master problem (using the simplex method) whereas the other 95% of the time is used in solving the sub-problems (generating the new columns using the dynamic programming algorithm).(求解子问题花费的时间远远大于求解RMP）

2021/11/21

• Utilizing individual picker skills to improve order batching in a warehouse
• doi: 10.1016/j.ejor.2017.05.002

Order picking is the most important, and the most expensive process in distribution
centers. It is estimated that order picking operations are responsible for 55% (Drury, 1988)
to 65% (Coyle et al., 1996) of the total cost to operate a distribution center.

Data on individual order pickers, readily available in many warehouses, can be used to improve the efficiency of the order-picking process.

Due to differences in picking skills,
differences in performance will exist among pickers.

capability (the degree to which a person is able to carry out a certain task, regardless of the time it takes) and skills (the speed at which such a task can be carried out)

WMS data of a picked order in a batch typically include: picker ID, roll cage IDs in which the items are picked, drop-off locations of the roll cages, time stamp of each order line, slot address per line, item IDs, and number of units picked.

we select the number of pick lines in a batch, total batch
travel distance, total pick item mass and volume in a batch, as well as the mean pick height level at which items are picked during the picking tour.

The problem is defined as follows. The orders that are to be picked during a day arrive early each morning.

1. joint order batching and generalized assignment (BatchGAP) model
2. routing pickers->batching orders->scheduling pickers->avoiding cogestion
3. Forecasting Expected Batch Execution Time
4. warehouse management systems (WMS)
5. skills:

• agility: ability to step on and off the pick truck ; modeled by the number of lines;
• driving skill; modeled by the batch travel distance;
• skill in picking heavy items (strength); modeled by the of total batch mass;
• skill in picking at low or high level, which may be influenced by picker height; modeled by the mean pick level of a batch;
• skill in picking large volume batches; modeled by the total batch volume.

6. multilevel analysis 多层次回归分析

2021-08-28：SMD，static move descriptor data structures

• 缩减vrp local search计算复杂度的技巧
• evaluating solution neighborhoods：对邻域解目标函数的计算，两方面，1降低单次evaluate的时间复杂度2减少需要计算的次数（没必要算的邻域解就不算了）
• when a local search operator is applied to a given solution, only a limited part of this solution is modified. Thus, to exhaustively examine the
  neighborhood of the new solution, only the tentative moves that refer to the modified solution part
  have to be evaluated.每次执行operator，解只更新了一小部分，所以之前计算过的很多邻域解并没有变化，没必要全部重新计算。
• metaheuristic strategies aim at intelligently guiding the local search process towards diverse trajectories of the solution space in order to escape from premature local optima and obtain high quality solutions.元启发式算法主要是增加了跳出局部最优解的机制，虽然简单，但是却是划时代的idea
• These move descriptors are stored into special priority queue structures which provide constant
  time minimum-retrieval and insertion, and logarithmic time
  update capabilities。smd用优先队列进行存储
• simple penalization mechanism specially designed for the VRP, which takes advantage of the static move descriptor entities, and is aimed at diversifying the search process.对不可行解进行惩罚，是提升算法搜索能力的重要一环，也就是允许搜索过程中当前解不可能，但是通过增加惩罚，让解不会不可行到离谱，类似拉格朗日乘子法；考虑解空间的话，会让你的解空间更加平滑，更不易陷入局部最优解

1. candidate list strategies, which generate
   only a small subset of all tentative local search moves
2. granular neighborhoods. These neighborhoods do not contain
   moves leading to features not likely to belong to high quality
   solutions, and are dynamically adjusted by exploiting information
   collected during the search process.

two-dimensional irregular strip packing problem

2021-04-30:periodic optimization for DPDPTW

• A periodic optimization approach to dynamic pickup and delivery problems with time windows
• doi:10.1007/s10951-020-00650-x
• (instance, feasible function, objective, {Min,Max})
  
• static optimization problem ——> dynamic optimization problem
• 
• 

2021/04/21

• Digital Twin Service towards Smart Manufacturing
• doi：10.1016/j.procir.2018.03.103
• digital twin is composed of three parts: physical products, virtual products and the connections between them.
• digital twin的三个level
  1. unit level: equipment
  2. system level: smart production line composed by machine tools, robot arms
  3. SoS(system of system) level: many unit and system level
• digital twin的三个流程
  1. The basic attributes, real-time status and other data are transmitted to the virtual models to drive the simulation and prediction.
  2. Then, the parameters of the virtual models are fed back to optimize physical entities.
  3. In the closedloop interaction process, the physical entities and virtual models co-evolute.
• three-dimensional structure of digital twin is extended to five-dimension, which are physical entities, virtual models, services, fusion data, and the connections among them.

1. Physical entities are the set of objective entities, which have specific functions to complete manufacturing tasks according to inputs and outputs.
2. Virtual models are the digital images of the physical entities, which can completely and truly reflect the lifecycle of the physical entities.
3. Services integrate various functions such as management,
   control and optimization, to provide application services
   according to the requirements.
4. Fusion data is the core driver of the digital twin, including the data from physical entities, virtual models and service, as well as their fusion data.
5. The connections among them connect the above four parts in pairs, ensuring real-time interaction and iterative optimization.

• PO(physical object) = {Basic, QoS, Cap, Status, Input, Output}
  1. Basic = {type, name, ID, address, color, delivery_time}
  2. QoS (quality of service) = {cost, time, reliability, trust, Fun_similarity}
  3. Cap = {shape_cap, precision_cap, parts_cap,size_cap}
  4. Status = {load, health}
• VM = {Ori_phy, Creator, Ori_ID, Cur_ID, Owner, QoS, Online_site, Input, Output}
• Data = {D_prov, D_source,D_ID,D_type,D_abstract}

2021/03/23

• RAWSim-O:(Robotic Automatic Warehouse Simulation (for) Optimization)agent-based and event-driven simulation framework

• manual picker-to-parts warehousing systems:人到货位仓库系统

• automated parts-to-picker warehousing systems：货架到人仓库系统

• Order picking – the process of retrieving products from storage (or buffer areas) in response to a specific customer request – is the most labour-intensive operation in warehouses with man- ual systems, and a very capital-intensive operation in warehouses with automated systems

• Achieving transportation economies (e.g. com- bine shipment, full-container load).
  • Achieving production economies (e.g. make-to- stock production policy).
  • Taking advantage of quality purchase discounts and forward buys.
  • Supporting the firm’s customer service policies. • Meeting changing market conditions and uncer- tainties (e.g. seasonality, demand fluctuations,
  competition).
  • Overcoming the time and space differences that
  exist between producers and customers.
  • Accomplishing least total cost logistics commen-
  surate with a desired level of customer service.
  • Supporting the just-in-time programs of suppliers
  and customers.
  • Providing customers with a mix of products
  instead of a single product on each order (i.e.
  consolidation).
  • Providing temporary storage of material to be
  disposed or recycled (i.e. reverse logistics).
  • Providing a buffer location for trans-shipments
  (i.e. direct delivery, cross-docking).

• The main warehouse activities include: receiving, transfer and put away, order picking/selection, accumulation/sortation, cross-docking, and shipping.

1. retrieving：入库作业includes the unloading of products from the transport carrier, updating the inventory record, inspection to find if there is any quantity or quality inconsistency.
2. Transfer and put away：the transfer of incoming products to storage locations.It may also include repackaging (e.g. full pallets to cases, or standardised bins), and physical movements (from the receiving docks to different functional areas, between these areas, from these areas to the shipping docks)
3. order picking/selection：the process of obtaining a right amount of the right products for a set of customer orders.
4. • accumulation/sortation*：picked orders into individual (customer) orders is a necessary activity if the orders have been picked in batches
5. Cross-docking ：is performed when the received products are transferred directly to the shipping docks (short stays or services may be required but little or no order picking is needed).

穷举搜索、贪婪搜索、束搜索

2021/02/22

• 关键链项目调度问题
• 资源约束项目调度问题
• 约束集理论：：（1）约束条件在所有系统中均
  普遍存在。如果有一个生产系统不存在任何约束，那么就可以无限降低成本
  或者提高产出，这在实际生活中是不可能存在的。实际上，任何妨碍生产效
  率提高的阻碍因素，都构成一个约束。约束理论应用于项目管理之中，就是
  把管理项目看成一个系统，妨碍项目提高效率和质量、降低成本的各种因素，
  其实就是该项目的约束条件。（2）约束存在表明项目存在改进的机会。
  关键链项目管理方法(CCPM)基于“局部最优的总和并
  不能导致全局最优”的约束理论思想，将项目管理重点放在系统的制约因素
  上，通过改进制约因素达到最大的有效产出

2021/02/08

A simulated annealing with a new neighborhood structure based algorithm for high school timetabling problems

• doi:10.1016/j.ejor.2009.09.014
• high school timeabling problems高校排课问题
• timeabling problems 排程问题
• conflict-free flexible timetables 首先追求一个可行解，让各方都满意，最后再制定合适的目标函数，持续优化

Timetabling consists in identifying an ‘‘optimal” allocation of a given set of events (courses, exams, surgeries, sport events) and resources (teachers, exam proctors, nurses, medical doctors) over
space (classrooms, operating rooms, sport fields) and time.

• 这句话提炼的好，事件+资源+空间+时间的一个组合，一个solution

Timetabling constraints can be classified into hard and soft constraints.

• 部分必须符合的硬约束，更多来自于物理属性，比如人不能分身；部分可以适当调整的软约束，大多来自人的主观因素，比如疲劳，课程效果等等，存在negotiation的空间。

2021/01/27

Ed Klotz a, Alexandra M. Newmanb,∗，（2013）Practical guidelines for solving difficult linear programs

• DOI:10.1016/j.sorms.2012.11.001
• 前两天碰到与调用求解器相关的问题，然后找到这篇，来学习一下

基本概念

• slack variable
• excess variable
• adjacent solution

好的句子

• Models corresponding to practical applications typically contain sparse A matrices in which
  more than 99% of the entries are zero.
• primal问题不好求，可以尝试求dual问题
• exploits the linearity of the objective and convexity of the feasible region in (PLP) to efficiently move along a sequence of extreme points until an optimal extreme-point solution is found. 凸性，局部最优就是全局最优
• (i) the algorithm discovers
  a ray along which it can move infinitely far (to improve the
  objective) while still remaining feasible, in which case the problem
  is unbounded, or (ii) the algorithm discovers an extreme-point
  solution with an objective function value at least as good as that of
  any adjacent extreme-point solution, in which case that extreme
  point can be declared an optimal solution.
• The solution time for simplex algorithm iterations is more
  heavily influenced by the number of constraints than by the
  number of variables.
• This is because the row-based calculations in the simplex algorithm involving the m × m basis matrix usually
  comprise a much larger percentage of the iteration run time than
  the column-based operations.
• The aspect ratio, n：m, generally indicates whether
  solving the primal or the dual problem, and/or solving the problem
  with the primal or the dual simplex method is likely to be more
  efficient.

2021/01/20

Patrizia Ribino*, Massimo Cossentino, Carmelo Lodato and Salvatore Lopes，（2017）Agent-based simulation study for improving logistic warehouse performance

• DOI:10.1057/s41273-017-0055-z
• 最近在用离散事件仿真+多智能体仿真做一些仓库里的仿真项目，但是在多智能体系统的设计上总感觉差点什么，所以还是返回来刷刷论文找找灵感。

基本概念

• agent-based simulation：基于智能体的仿真
• Logistic warehouses：物流仓库
• unproductive bottlenecks：无生产效率的瓶颈
• automatic guided vehicles：AGV，自动导引小车

摘要和introduction

• 看完摘要，发现本文是要讲使用AGV的自动化仓库如何通过基于智能体的仿真方法来提高效率，发现瓶颈。

• Simulations can be used in different contexts for evaluating the behavior of complex systems,
  where conducting simulations allows alternative decisions to be examined. The effects of these alternatives can be tested without conducting experiments in a real environment, which
  is often prohibitive in terms of cost or completely unfeasible. In many practical fields, simulation is the most affordable way of understanding how the numerous variables interact and constrain the performance of systems.

• Warehouse

  1. a set of gates where lorries are parked to unload their cargo;大门
  2. a set of recharging areas where AGVs recharge their batteries充电区
  3. sorting area where goods are processed and forwarded to a new destination;分拣区域
  4. a buffer area where unidentified goods are placed temporarily;缓冲区
  5. a set of AGVs
  6. a set of waypoints, which are landmarks for AGV navigation;路径点
  7. a set of optical paths for AGV navigation, where each path is represented by a set of segments

• Agents are autonomous entities that can work independently and/or cooperate with other agents in order to achieve some goals.

  1. ABMs are essentially decentralized:In contrast to system dynamics or discrete-event models, there is no place to define the global system behavior in an ABM. knowledge of the global interdependencies。无法知晓系统的全局行为或者群体行为，通过仿真智能体的个体行为，individual behavior
  2. In ABMs, there is an intuitive ontological correspondence between agents and the real-world actors.ABMs can incorporate discrete-event simulation mechanisms。多智能体仿真可以和离散事件仿真相结合。
  3. ABMs have great flexibility and efficiency when modeling different types of systems.
  4. ABMs are well suited to visualization适合于可视化a powerful method for verifying, validating, and explaining the model.
  5. ABMs are usually easier to maintain, where model refinement generally results in very local (not global) changes.

2020/4/30

Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan, (2002)A Fast and Elitist Multiobjective Genetic Algorithm:NSGA-II. IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION

• DOI:10.1109/4235.996017
• 最近发现很多人做的问题是多目标优化，虽然最简单的处理方式就是通过线性加权进行线性化处理，但是各目标的权重难以给出，因此直接求解帕累托最优解集合，然后全部给出让决策人员自己去根据自己的偏好再进行选择是不错的思路。

基本概念

• Pareto-optimal solution：帕累托最优解
• Dominate&Dominated：支配（完全占优）和被支配（被完全占优）
• Non-dominated solution：非劣解
  帕累托最优解，是从经济学那边借来的概念，主要指在多个目标评价指标的情况下，公说公有理婆说婆有理。比如，现在对男人进行排序，以高富帅三个指标为准。你说吴彦祖和我比，那吴彦祖确实不论身高长相和money都比我强，输给他我心服口服。此时，吴彦祖就dominate了我，我被吴彦祖dominated了。此时就会出现一个可怕的事实，如果世界上只剩下了我和吴彦祖两个男人，而在只能2选1的情况下，只以高富帅三个标准，没有一个人会选我。所以在此时我的对偶价格其实是0，没有价值，完全可以销毁。（那为啥我还有活着的意义呢？是因为吴彦祖一个人不够用了，吴彦祖无法完成所有的工作，在比较优势下他只能选择做收益最高的工作，而我可捡他剩下的，当然里面主要还有一个隐性条件，时间，吴彦祖无法克服时间的维度，如果他会影分身，我就真的凉了）此时我叫来了我的爸爸马云，替我与吴彦祖进行比较，吴彦祖比马爸爸帅，胜1，吴彦祖比马爸爸高，胜2，吴彦祖比马爸爸穷，负1。好的，吴彦祖赢了！！！错。我选马爸爸。因为在不同目标上，吴彦祖与马云各有胜负，因此他两互不完全占优，因此是两个非裂解，但是不一定是帕累托最优解，因为如果存在比吴彦祖高帅比马云有钱的人存在的话，吴彦祖和马云就同时被支配了。但是我们假设不存在这样的变态，就可以任务，马爸爸和吴彦祖是帕累托最优解集合中的两个帕累托最优解。

本文目的

本文的目的是为了求解多目标优化问题，尽可能的寻找出所有的帕累托解集合。但是我们学了那么多的单目标优化算法，都是针对一个点（解）的寻找，因此需要对这些算法进行一些改进。既然是寻找一组帕累托解，我们自然会想到，群智能算法是不是天生一对。其实，遗传算法，粒子群算法等等都有被用来求解多目标优化问题。而本文就是再NSGA算法的基础上进行改进，提出了NSGA-II算法，是几位印度大佬的成果，引用两万五，真大佬，对印度人的印象改观了，人类大同！

• Pareto frontier：帕累托前沿，由帕累托解集合构成的超平面
• Elitist Nondominated Sorting Genetic Algorithm NSGA-II
• 精英策略Elitism，这个策略可以更加有效得保存较优解，维持种群的平均质量

Fast Nondominated Sorting Approach

内容

• Constrained multiobjective optimization：约束下的多目标优化，强调约束

Two common features on all three operators were the following: i) assigning fitness to population members based on nondominated sorting and ii) preserving diversity among solutions of the same nondominated front.

• nondominated sorting:非劣解排序 ：其实解决的一个最简单的问题，如何给种群中的染色体排序
• 保持种群中各类非劣解的密度均衡，就是控制种群的基因多样性，不要出现一家独大的现象，公平，公平还是他娘的公平。

2020/4/25

Bramde Jonge, Philip A.Scarf, (2020)A review on maintenance optimization.EJOR

• https://doi.org/10.1016/j.ejor.2019.09.047
• maintenance optimization：维修保养优化，maintenance是有维修保养的意思，侧重于在equipment彻底宕机前的，检修。通过合理、有效的维修保养，工业生产设备、载运工具、电力系统等等设备都能延长使用寿命、提高生产效率、降低意外风险等等，所以maintenance在生产生活中都起着重要的作用。其实在我们的生活中，最常见的场景就是对自己的私家车进行维修保养，你需要规划好哪些日子定期去4s店进行保养，在保养期间，你无法使用车辆，此外还需要选择去哪家店进行保养，进行什么级别的保养，保养的频次等等。延伸到其他大型系统中，这个问题就变得更加复杂，维修保养会打乱生产节奏，但是为了保证不出更大的乱子，你又不得不进行maintenance，所以存在tradeoff，也就有了运筹优化发光发热的余地。这篇问题是最新一期EJOR的综述文章包括了2001-2018年的一些好文章，看这篇文章的原因是最新一期ROADEF的赛题是电力系统maintenance优化问题，因此需要做一些前期调研。而且我这半年在家里蹲，一事无成，只能硬着头皮冲一下。

The exact features that studies on maintenance optimization consider are generally inspired by practical situations, but only a small number of studies in this review apply their model to a real life case.

• 这句话就有意思了，说道虽然很多问题的问题是来源于实际情况，但是他们的研究成果很少很少被应用在实际中，这很正常，一是说明学术研究还不足以处理现实中复杂的问题以及不确定性带来的扰动，二是说明业界受限于基础设施的不足，无法采用OR的一些优化方法还是脑袋拍起来更爽快，三是企业必然藏着掖着了这是人家的商业机密，当然论文太难写也是原因。但是文章也说了具体的原因:

1. models must be interpreted
2. decision support systems need to be developed
3. data are required
   这么看，说明信息化、数据化和智能化之路道阻且长

Returning to our review, it is our purpose that it is useful for both researchers and practitioners. Researchers can use our review to identify studies that are related to their problem, to position their study in the literature, and to get inspiration for future research. Practitioners may use our review to provide guidance on the identification of models that most closely match their maintenance problem.
这段说明了本文的目的，我觉得说的非常好，将researchers与practitioner之间的gap烫平。

• maintenance management 作为 asset management的一部分，asset management关注设备的整个生命周期，而maintenance management主要关心operational phase of assets

1. agent: a stakeholder in the modeling process
2. system:is an asset that performs an operational function single and multi
3. deterioration state space: two states, three states, a discrete state space, and a continuous state space, and devote subsections to each of these categories.
   • two state space: functioning or failed 好的或坏的
   • three states： functioning or failed 好的或坏的 + 介于好坏之间的状态（一般是性能变差，成本变高等等）delay-time model:good, defective and failed
   • discrete state space have a countable, generally finite number of states
   • continuous state space the deterioration level can take any value within a particular interval
     状态空间肯定无法完全trace，最好提炼出几个典型的状态，便于模型建立与求解，但是需要很好地识别。
4. actions preventive and corrective 预防性检修和修理式检修 CBM condition based maintenance

2020/3/01

Alberto Fernández, Holger Billhardt, Sascha Ossowski and Óscar Sánchez,(2020)Bike3S: A tool for bike sharing systems simulation.Journal of Simulation(JOC)

• https://doi.org/10.1080/17477778.2020.1718022
• vehicle sharing system载具共享系统 共享单车，共享汽车

1. the composition of the fleet
2. the distribution of vehicles
3. rental stations 有桩无桩，摩拜ofo是无桩的，政府搞的是有桩的，大家肯定用无桩的，但是估计这篇要研究有桩的了，没劲。有好几个原因，一是用户体验，无桩想停哪就停哪，而有桩只能在一些地方设置站点。此外，站点的slots有限，说不定你到一个地方，发现没有车位了，白瞎，换不了车。因为OD之间的流量是不均衡的，所以百分百会出现这种情况。而无桩会导致乱停乱放也需要rebalance，需要更多的人力去管理以及更合理的奖罚措施驱动用户去更好的完成停车操作，然后还考验素质和城市管理水平。

Several research works have been done so as to optimise the use of trucks to keep the bike sharing fleet as balanced as possible, either statically (typically at night)or dynamically (during operation) trying to match the expected demand with the available resources.

• https://github.com/gia-urjc/Bike3S-Simulator
  

2020/2/15

James F. Campbell，Morton E. O’Kelly,(2012)Twenty-Five Years of Hub Location Research.TS

• http://dx.doi.org/10.1287/trsc.1120.0410
• Hub location problems involve the location of hub facilities through which flows (e.g., of passengers or freight) are to be routed from origins to destinations (e.g., cities).
• The main focus is the cost of travel on these links—and opportunities to exploit transportation economies of scale.

2020/2/14

Mobina Masaeli，Sibel A. Alumur∗, James H. Bookbinder,(2018)Shipment scheduling in hub location problems.TRB

• https://doi.org/10.1016/j.trb.2018.07.003
• 本文是老板发我的，用来解决快递公司中转场的选址和网络转运设计，当然，核心是hub location problem HLP，我们知道，HLP主要是解决location和allocation两个interactive的决策。但是这样往往不足，因为简单的OD运输需求与现实不符，未考虑现在时效产品的存在，而且公路运输中，车辆较小，是离散的规模经济函数。因此有很多想象的空间。需要解决的决策，包括，班次衔接，货物囤积等等。

we incorporate shipment scheduling decisions into hub location problems. Our aim is to determine the optimal locations of hubs, hub network structure, and the number of dispatches to operate on the hub network as well as the time period of dispatching each vehicle from a hub.

• hub facilities: unloading, sorting, loading 中转场，集货-分拨-散货，负责卸货，分拣，装货，目的是把不同始发地来的货物中的同一目的地的货物聚拢在一起，装在一辆车上，形成规模效应。因此，选择在哪里设置hub，多大规模的hub很重要
• hub-and-spoke network:轴辐式网络，多级中转场。
• 运输在h&s网络中的货物以OD为特征，例如北京到上海。因此我们可以定义出某个货物 P（O,D,weight,volume,time-stamps{timeO,…,timeD}, 时效）
• single allocation：spoke点只能分配的看一个hub
• 时效产品：same-day当日达，next-day 次日达
• shipment scheduling 运输计划，spoke-hub/hub-hub/hub-spoke什么时刻发车，多大的车，哪些货物上这辆车

问题分析

• single-allocation hub location and shipment scheduling problem
• N ： demand nodes set
• H ： potential candidate hub locations set
• T ： planning horizon 离散化到了|T|个time periods t=1,…,|T| time period = [0,1) T = 3 [0,1),[1,2),[2,3) 三个时间段
• 这里只考虑了inter-hub的shipment scheduling，因为spoke到hub的车都是满车，到点就发出来了，而hub-spoke的时间只要在时效前就行了，可以倒推。
• Each vehicle traverses an inter-hub link from k ∈ H to l ∈ H in transit time τ kl. Hence, flow entering an inter-hub link (k, l) at time t ∈ T, arrives at l ∈ H at time t + τkl. 主要是增加了一个时间维度
• 这里不好处理的点在于，网络中的货流是动态的，不像传统HLP中，参数都是给出，直接计算。本文，增加了一个planning horizon的概念，并且离散化，对时间进行离散化，这个技巧很好用，也很常用，在咱们这个系统中，本来就可以用离散事件仿真来描述，所以对时间进行离散，比较合理。

2020/2/13

Zhenzhen Zhang, Zhixing Luo, Hu Qin, Andrew Lim, (2019) Exact Algorithms for the Vehicle Routing Problem with Time Windows and Combinatorial Auction. Transportation Science

• https://doi.org/10.1287/trsc.2018.0835
• 这是从通过秦虎老师看到的TS上的一篇论文，提出来VRPTWCA这个新的VRP的变种问题，考虑了允许将部分需求外包给第三方物流公司3PL的情况。实际的意义在于，例如顺丰是有部分自有车辆的，自有车辆的优点是，方便管理，稳定可靠，但是如果运输任务不足，成本会比较高。而第三方车辆通过平台竞价，获得某班次的运输任务，灵活自由，可以弥补货运任务的波动带来的不确定性。但是第三方车辆通常难以管理约束，经常出现晚到，违约的情况，且车型不一，较难管理。因此最简单的办法就是把货运需求分割：
• 总货运需求D = 基本稳定货运需求A + 浮动峰值货运需求B ，其中自有车辆负责A，第三方车辆负责B。当然，什么样的比例需要根据实际情况去区分，且自有车辆闲暇时也可接其他货运需求。但是还是一句话，业务货量足够多，就不怕多投资，开源节流，开源最重要。

问题分析

• 这个问题的场景在于，部分客户点，直接外包给3PL，这样的话，自有车辆的路径规划就是简单的VRP了。但是显而易见，这样就增加了许多决策变量：将哪些客户点外包给3PL。当然，以什么样的价格，外包给哪家3PL，都可以成为决策。而OR问题的基本准则是，找到合适的决策变量，满足约束，实现目标。而将部分决策固定下来，作为parameters输入值，就会将问题简化，便于处理。
• 可以复盘，哪天领导给了你一个简单的VRPTW求解，共50个点，你觉得太简单了。但是，其实领导已经将其他550个点外包出去了。你想知道他怎么决策的，为啥就把那550个点外包了。因为昨晚他喝得挺爽，就这么决定了，毕竟咱们自己的车队也就能处理50个点左右。这时你主动请缨，为领导分忧解难，说，让我来，我能解决在各公司给的报价下，咱们自个车辆运输哪50个点的问题，这样就不用喝酒了，用数据说话，量化指标。这就是本文的重要价值。拓展了，困难了，但也变强了。

模型分析

• 文章建立 arc-flow和set-partitioning两个模型，模型倒没啥特别的，主要在于，部分客户点用Yu决策变量执行，u是一个bid，包括a set of customers,和一个price Pu，每个客户点要么被自有车辆访问，要么被外包出去。当然，外包方案中也存在相互冲突，每个客户点只能访问一次。
• 这个问题挺好，用另一种思考方式就很好理解，如果每个bid只能包含一个客户点，有一个成本，就可以看做，允许部分点我不服务，不服务的点有一个惩罚成本，使得整个的成本最低。与VRP with profit挺像的，都是允许customers的selection。不同之处在于，3PL自己也根据多个点的特征，提出了group报价。这个group报价一般比单个报价要便宜，规模经济呗。挺好的，很有意思。
• 这个模型与航班恢复里也很像，客户点要么被route执行，要么cancel，但是这里还有个cancel组合的概念，成本不一。
• 因此很容易想到，上列生成。使用Branch&Price&Cut就更好了，更快更强。

Valid Inequalities 有效不等式

• 有效不等式用来放入主问题模型，作为cut切割解空间，加速模型求解。这些cut作为约束，通过dual price传递给子问题。

1. k-path inequalities
   • subset S in N ks = the minimum number of vehicles required to serve all the customers in S
   • 其实就是，一组客户点，至少要几辆车才能服务完。通常由容量决定，时间窗与他们之间的距离也会影响
2. Capacity Inequalities

算法求解之Branch-Price-Cut

Initial Feasible Solution

• 好的初始解，可以加快算例求解，在最好的情况下，初始解==最优解，一次直接全垒打。
• 设计了一个backtracking heuristic算法生产初始解。先将bids按客户数量从高到低排列，贪婪得选择尽可能多的compatible的bids，这样就将客户点规模缩小了很多，剩下的点就按照简单的VRPTW生成路径，使用*Potvin and Rousseau(1993)*的 regret insertion继续求解。

Pricing Problem

• 求解子问题的第一步是先把 子问题的目标函数写出来，也就是路径成+各个约束的dual value

• 此处得到的子问题是ESPPRC 不可重复的带资源最短路问题

• 这里选择的算法还是 Label-Setting Algorithm

• The label-setting algorithm is a type of dynamic programming algorithm that enumerates all nondominated routes by state propagation.

• A state is a partial route from node 0 to any node i in V`,represented by a label Li = (Ci *reduced cost) (Ti earliest time) (Di accumulated demand) (Mi,s the numer of nodes in S that have been visited by

• 下面是一些加速方法

1. t bounded bidirectional search, which divides the extension of the label-setting algorithm into the forward extension and the backward extension according to a chosen critical resource.双向标号法
2. ng-route relaxation
3. q-route relaxation

Tabu Search Column Generator

• 只要能迅速找到reduced cost为负的route，不论用什么方法都行。因为标号法还是太费时间了。
• 所以采用heuristic去搜索是可行的，因此文章使用了基于禁忌搜索的column生成模块

In the literature, one common way to speed up the convergence of column generation is to use a heuristic to find columns with negative reduced cost first,and to invoke the label-setting algorithm only after the heuristic fails (Desaulniers, Lessard, and Hadjar 2008;
Archetti, Bouchard, and Desaulniers 2011; Luo, Qin,and Lim 2014).

• 在feasibility test这块，还是采用了Vidal大神的segment-based 的方法，比较是O(1)的时间

Cutting Separation Algorithms

• 这部分是Cut里面的核心，如何生成新的cut constraints，通常也是使用启发式算法