Analysis of the Expected Possession Value in RoboCupSoccer Simulation 2D

Abstract

RoboCupSoccer Simulation 2D(RSS2D) simulates a soccer match between two teams of 11 autonomous mobile robots each on a virtual two-dimensional plane. In soccer, the ball and players are always on the move, leading to a constantly changing game situation. RSS2D agents thus need to constantly analyze the match situation and predict the actions of enemy agents in determining their actions. However, it is not easy to analyze a player’s actions, because it is necessary to combine specialized knowledge with consideration of the intentions of all players’ actions. In human soccer, an analysis method using the expected possession value (EPV) has been proposed to solve this problem. The EPV is a measure of the probability that a team will score or concede a goal according to the game conditions and player actions. In this study, we apply the EPV to RSS2D to analyze match situations and agent actions. We also examine whether EPV-based analysis is effective in an RSS2D environment. We create a dataset by extracting agent actions and ball position information from RSS2D log files. Next, using the dataset as input, we train several estimation models using a convolutional neural network. Then, by combining the multiple estimation models, we create an estimation model for the EPV of passing. The effectiveness of the estimation model for the EPV of passing is confirmed by analyzing game conditions and agent actions.