Using Multivariate Adaptive Regression Splines in the Construction of Simulated Soccer Team's Behavior Models

• Main Authors: Pedro Henriques Abreu, Daniel Castro Silva, João Mendes-Moreira, Luís Paulo Reis, Júlio Garganta
• Format: Article
• Language: English
• Published: Atlantis Press 2013-09-01
• Series: International Journal of Computational Intelligence Systems
• Subjects: Knowledge Discovery from Historical Data; Data Mining; Feature Selection; Soccer Simulation
• Online Access: https://www.atlantis-press.com/article/25868429.pdf

In soccer, like in other collective sports, although players try to hide their strategy, it is always possible, with a careful analysis, to detect it and to construct a model that characterizes their behavior throughout the game phases. These findings are extremely relevant for a soccer coach, in order not only to evaluate the performance of his athletes, but also for the construction of the opponent team model for the next match. During a soccer match, due to the presence of a complex set of intercorrelated variables, the detection of a small set of factors that directly influence the final result becomes almost an impossible task for a human being. In consequence of that, a huge number of software packages for analysis capable of calculating a vast set of game statistics appeared over the years. However, all of them need a soccer expert in order to interpret the produced data and select which are the most relevant variables. Having as a base a set of statistics extracted from the RoboCup 2D Simulation League log files and using a multivariable analysis, the aim of this research project is to identify which are the variables that most influence the final game result and create prediction models capable of automatically detecting soccer team behaviors. For those purposes, more than two hundred games (from 2006-2009 competition years) were analyzed according to a set of variables defined by a soccer experts board, and using the MARS and RReliefF algorithms. The obtained results show that the MARS algorithm presents a lower error value, when compared to RReliefF (from a pairwire t-test for a significance level of 5%). The p-value for this test was 2.2e-16 which means these two techniques present a significant statistical difference for this data. In the future, this work will be used in an offline analysis module, with the goal of detecting which is the team strategy that will maximize the final game result against a specific opponent.