运动科学综述阅读(3): Team Selection

Role of meta-heuristics in sports

总结：
选择出场队伍的问题。 一种权衡的艺术，因为大部分可选项之间的差异都不是绝对的，比如方案A可能在进攻上优于B，但是防守上若于B；或者A能力不是特别高，但是价格低很多… 所以需要考虑很多条件；而考虑的因素越多，整个问题就越复杂，导致更难解决。
算法方面： 大部分这方面的论文都采用GA或者变体（NSGA II），有一篇[Prakash(2016)]用到了ML+Meta-heuristic+Memetic Algorithm.
解决问题的考虑角度： 前期使用单目标优化，后期倾向于多目标。 目标/适应度函数是很关键又很难的问题。

下一节是line-up问题，Cambridge Dictionary的一个解释是the order in which the players in a baseball team hit the ball。 可能是兵对兵、将对将的问题？ 期待:)

论文内容：

Team selection

external factors; raise more questions;
earlier studies: integer program, analytical hierarchical
Brettenny et al., 2012
Kamble et al., 2011

Summary Table

genetic algorithm and memetic algorithm

factors taken into account in implementing suitable performance measures
only focus on cricket and basketball. For general discussion on those sports, see
(Skinner and Goldman, 2015,
 Swartz, 2017)

P.S. 其实大部分都是cricket

Omkar and Verma (2003):

fitness function
● GA fitness measure is defined for each of these criteria for each player and the fitness of each player is calculated as the sum of these fitness values. The fitness of the team is the addition of fitness of each player alogn with a constraint over the composition … In addition, data regarding the past performance…have been considered in the study.
●permutation presentation is used where each gene in the string represents a number of the player in the team
○not enough providing evidence on the competency of the method; selecting 15 players out of 30 cannot be considered as an easy task, and the need of a meta-heuristic is emphasized here
○worth investigating more on the fitness function used in the future

Sathya and Jamal, 2009: GA detailed descriptions on the formation of the fitness function.

●some results
●no detailed comparison or a proof
●Both have not paid any attention to the performances of the opposing team

multi-objective optimization problem.

NSGA II algorithm Deb et al. (2000) Ahmed et al., 2011b
team selection -> bi-objective problem

Pareto front: An individual A is member of the Pareto front if there is no individual B in the population that is at least as good as A concerning every criteria and better than A in at least one criteria. 即不存在B在每个方面都比A好/相同。

knee point

Ahmed et al. (2011a) : extends Ahmed et al. (2011b) by adding more objectives; changes the chromosome representation
●knee region approach -> selecting suitable solutions from the Pareto front
●reveal the importance of using a multi-objective optimization over a single objective for a better team selection
●extended version is presented in Ahmed et al. (2013) with more emphasis on a meaningful representation schema.

Burney et al. (2012)

●generalized approch, quantitative measures with new measures
●four different types of players
●fitness of player -> selection probability to the team
●average of fitness of each selected players -> fitness of team
●some inspiring results, while lack of considering the individual performance in fitness calculation

Prakash (2016)

●combining machine learning and meta-heuristics: random forests based recursive feature estimation algorithm => 用这种算法，确定performance indices
●memetic genetic
●binary string representation
Kumarasiri (2017)
give some insight about the importance features, such as qualitative features which are harder to consider but might be useful

Gamed related matrices:
●Performance Index Rate(PIR)

The teams found by NSGA II have a theoretical high level of performace at a low cost in comparison with the teams playing in the ACB League the following season. Further, it enlightens and inspires the reader to use it for other sports.