运动科学综述阅读(6): Equipment,Scheduling and Discussion on meta-heuristics

本次把最后一部分读完了，在算法的应用领域上，设计Sport Equipment Optimization, Scheduling and Ranking， 以及最后讨论了Meta-heuristics的应用和前景及困难

Data mining for predictions in sports

video analytics in sports

EA

Tsakonas et al. (2002)

football winner predictor model

• GA fuzzy logic based prediction model
  GA->optimal weight adjustment for fuzzy model learning phase
• GP
  genetic operators: tournament selection with node mutation and shrink mutation
  fitness-> correct score ratio
• NN
  GP outperformeed

Khemka et al. (2005)

football
leg and foot movements for kicking a soccer ball
focused on finding optimal parameter set for muscle control with the objective of maximizing forward ball speed
PSO, ES

Rotshtein et al. (2005)

fuzzy model, tuned by GA and NN
football, predictions based on the previous game results of the two teams

Zhao (2008)

soccer video event retrieval with multi-modal features
GA( interactive GA, where fitness is assigned by a user subjective evaluation(Takagi, 2001))
chromosome: 8 audio visual features
one point crossover genetic operator
select second generation: minimum euclidean distances between chromosomes and targets to videos

• At the end of iterative process, the user preference video was obtained
• human evaluation can cause bias, making the usage difficult in the real world scenarios

Narasimhan et al. (2010)

GA

• shots and events: a shot is a group of segmented consecutive frames having similar visual content. A sequence of shots generates an event
• label: based on commentaries available in cricket websites as ‘six’, ‘four’ ‘out’ and ‘run’
• fitness: linear fitness measured, consider three measures: interests, variety, and spread
• stochastic uniform selection, scattered crossover, adaptive feasible mutation GA operators
• result: At the end of a run, the GA was able to select the events to be included in the summary according to the importance level

Li and Wang (2011)

GA with BP neural networks to the problem of EEG feature selection and pattern classification
EEG: Electro-encephalo-graphy脑电波图 中间应该是没有连字符的，居然这么长的一个单词
find the affectable characteristics of long-term exercise training on EEG

• 5 brain waves-> tested for impact
• 19 features-> binary string, having crossover and mutation probabilities of 0.5 and 0.05 respectively
• classified data using BPNN fed as input to GA, in order to select the optimal feature subset
• BPNN find the optimal features that can distinguish motor imagery EEG of these specialties
• revealed that long sports training have a great impact on the activity pattern of β and α brain waves

few sports related studies that were benefited by SI algorithms applied for data mining tasks

SI

Wang et al. (2014)

improved AFSA: initializations were done with the help of chaotic search
optimize performance in the game of table tennis
the heuristic initializations, although discourage in normal meta-heuristic practices, can work well with some kind of problems
取得更好地优化，因为挖掘到更多的特征

Behravan et al. (2019)

PSO
find roles(positions) of football players based on clustering
代替了cluster中的KNN的方法
goodness: Calinski-Harabasz index
simulations were carried on both synthetic data and on real world football data
Bose and Chakraborty (2019)
cricket
Cumulative Sum (CUSUM) control charts, which depends on the parameters and the parameters set are optimized by GA
accelerating the run rate

Shingrakhia and Patel (2020)

latest SI to optimize event recognition and summarization in the game of cricket
improve Narasimhan et al. (2010)

• by collecting both event driven and excitement based features
  summarizing a full-length video while preserving the important moments
  Deep Neural Network + EPO -> EPO to find optimum weight parameters; DNN find semantic features accurately based on the optimized parameters
  applied a meta-heuristic in an indirect way to boost the data mining process

Nelikanti et al. (2021)

SSO -> SMO + SSA =>for training a deep Long Short-Term Memory classifier (LSTM) for path prediction

• better decisions taken on leg before wicket situations by using the prediction paths
  wicket是cricket中“三柱门”，可能指比较接近被进球的情况 “为防止球击中三柱门，三柱门由击球手使用球板保护”（Wiki 三柱门）

one of the key research in sports that introduced a new algorithm rather than applying an existing algorithm

Summary

1.directly involved in the extraction or classification process
2.optimizer of other methods or models such as nn or fuzzy logic based classifiers

Sports equipment optimization

equipment have a direct influence on the performances of the players or the team(Shan, 2008)
implementation of optimized and player friendly sports equipment for better performance is an open research area where there exist many avenues to explore

Marcelin et al. (1995)

use GA to determine the optimal structural damping solutions for skis
dampness: ability to absorb the natural vibrations inherent in the interaction between the ski and the snow for a smooth ride
focus on finding the best positions for the dampers or the stacking sequence of composite materials in order to maximize a modal damping factor

• in some situations: with or without variable layer thickness
  m-encoded chromosomes -> stacking sequence
  partially-mapped crossover and inversion mutation with probabilities of 0.6 and 0.1
  reveal the comfort-ability and effectiveness of GA in finding solutions in the discrete domains

McHutchon et al. (2006)

Optimizing physical properties of hockey sticks
GA

• Power and the control of the stick are defined based on the mass of the stick
  no any clear definition of the fitness function
  … Although clear explanations are not present, as an early development, the effort of the study is admirable

Shimoyama et al. (2011)

GA
sport shoe sole - with an aim of improving its operability which allows an athlete easy to keep running with low energy consumption
Studies have proven that operability depends on shoe weight and sole stiffness

• chromosome: 11 design real value variables identified for the mid-sole structure
• two problem formulations
  ○single objective constraint optimization
  ■constraints based on the weight of the sole obtained from the prior studies
  ■Kriging model(Sacks et al., 1989) to construct a response surface, which act as fitness fuction in the form of a simple algebraic function
  ○multi objective optimization
  ■non-dominated solutions using the same fitness criteria

Mazloomi et al., 2020

cricket bat
Algobat with the theme of cricket for all, experiment by Univeristy of British Columbia
two features of bat:

• vibrational nodal points
• center of percussion(COP)
  optimize the distance between the two points which are responsible to rebound a ball with increased celocity via RMSD(root mean square of the distance).
  28 viriables
• GA to obtain the initial optimum values of the two way optimization process
• then fed to another optimization method MMA(method of moving asymptotes) for further minimization
  in both studies conducted on cricket and baseball bats, they have not prototyped and measured the performances

Mazloomi and Evans (2021)

wooden baseball bats + whether wood species (ash vs maple) have an influence on bat geometry
10 variables
GA
in both studies conducted on cricket and baseball bats, they have not prototyped and measured the performances

Scheduling and ranking

Cassady et al. (2005)

• degree of victory achieved by the winner
• relative distance between every pair of ranking positions
  GA+local search techniques-> near-optimal solutions
  integer chromosome-> position to identify the rank
  fitness-> use of degree vicotry
  special crossover and mutation operators for the quadratic assignment problem

Guangdong et al. (2007)

NP-hard
GA+ACO-> optimum schedule in which the competitors who taking part in two sequential games are minimized
GA->optimized schedules and update pheromone
ACO->obtain solutions applying crossover and mutation operators
benchmark traveling salesman problems

Connor et al. (2019)

GP
tram sports training plans
maybe a typo for team sports
specific fitness function related to the aim of the study

Connor et al. (2021)

effective training sessions/models for team sports
high performance on match days and minimize large fluctuations in the daily training load

难道不应该平时多练，比赛前休息吗？

fitness: 2 matrices: Velocity based running intensity(VRI) and acceleration based running intensity(ARI)
GP grammatical evolution, an extension to canonical genetic algorithm
integer flip mutation, fixed single point crossover

Summary

• many objectives and several constraints
• considering different sports such as cricket, soccer, basketball, and commonly sports tournaments, scheduling game seasons are addressed in several studies (Biajoli et al., 2004, Lim et al., 2006, Sch¨onberger et al., 2004, Willis and Terrill, 1994, Wright, 2006).
• Scheduling can be considered as one of the earliest problems addressed in operations research as well as in the field of meta-heuristics. Many of such research have used simulated annealing and memetic algorithm to solve the optimization task.

Summary

本综述阅读系列中提到的算法是hybrid approaches or new variants rather than applying the problem to an existing algorithm
notable works:

• Fister et al., 2018, 2015, Schaefer et al., 2015

Disscussion

meta heuristic is of popularity

time

rapid involvement during the last 5 years, further establishes the trustworthiness of extremely attractive abilities of nature inspired meta-heuristics

Algorithms

EAs(particularly ga) and SI(PSO) well verified on their performance, parameter tuning, different representation
-> attract sports community to use these two algorithms without an effort
SA the only physically inspired algorithms. Widely used in schedling tasks, 30+ years of history, well-established
available in built-in package for SA, GA, GP and PSO

Sport categories

cricket and soccer benefit more than others

• large state space due to higher number of players
• money
• public availability of historical data
  work on basketball not much famous-> less number of players, which limit the use of meta-heuristics(optimization)

attraction factors

• Availability of data
• importance of the outcomes to the betterment of the field

Role of meta-heuristics

fully or partially adopted to solve a particular task

• fully implemented
  • group selection or optimal alignment
  • treat as optimal optimization problem
• partial involvement
  • data mining or performance analysis
  • ML as major role, making such task success
  • meta- heuristics as booster, optimizing or adjusting the hyper parameters of ML algorithms

figure 10: Sports Vs Tasks categorization of the literature; Background colors represent different categories of algorithms: EA, SI and PI

Foreseeable chanllenges

optimal choice of algorithm for a particular task

• lack of knowledge on meta-heuristics in the field of sports(therefore tendency is to use ready-made software solutions)
• higher number of hyper parameters. Tuning requires expert knowledge.
• suitable fitness function, proper representation techniques
• expensive fitness functions for simulation tasks
• multiple object optimization for multiple tasks. Need to find solutions over Pareto fronts
  potential of optimization algorithms, revealed in NN and CUESUM(Section Data mining for prediction in sports)

main intentions of such review

• awareness of sport analytics groups on meta-heuristics
• discussion on the implantation of meta-heuristics in sports