Analyzing and understanding strategies applied by top soccer teams has always been in the focus of coaches, scouts, players, and other sports professionals. Although the game strategies can be quite complex, we focus on the offensive or defensive approaches that need to be adopted by the coach before or throughout the match. In order to build interpretable parameterizations of soccer decision making, we propose a batch gradient inverse reinforcement learning for modeling the teams` reward function in terms of offense or defense. Our conducted experiments on soccer logs made by Wyscout company on German Bundesliga reveal two important facts: the highest-ranked teams are planning strategically for offense and defense before the match with the largest weights on pre-match features; the lowest-ranked teams apply short-term planning with larger weights on in-match features.
© Copyright 2022 Machine Learning and Data Mining for Sports Analytics. MLSA 2021. Communications in Computer and Information Science. Published by Springer. All rights reserved.