Learning Opponent’s Strategies in the RoboCup Small Size League

One of the machine learning challenges posed by the robot soccer domain is to learn the opponents strategies. A team that may be able to do it efficiently may have the advantage to adapt its own strategy as a response to the opponent’s strategy. In this work, we propose a similarity function to compare two teams, and consequently their strategies, by the ability of one team to mimic the behavior of the other. The proposed function can be used to classify opponents as well as to decompose an unknown opponent as a combination of known opponents. We apply the proposed function to classify opponent’s defense strategies in real world data from the RoboCup Small Size League collected during the RoboCup 2007, RoboCup 2008 and USOpen 2009. We also use this similarity function to discover patterns in the logs of these championships, such as, similar teams and the number of major defense strategies.