[2M4-OS-37a] OS-37

[2M4-OS-37a-04] Tactical Decision Analysis in Soccer Using Reinforcement Learning with Interpretable Low-Dimensional States

〇Kenjiro Ide¹, Taiga Someya², Kohei Kawaguchi³, Keisuke Fujii^1,4 (1. Nagoya University, 2. The University of Tokyo, 3. The Hong Kong University of Science and Technology, 4. RIKEN)

Keywords:Multi-Agent, Time-Series Modelling, Sports

In sports like soccer, modeling and analyzing matches is challenging due to the continuous interactions among multiple players. Rule-based models struggle to capture all players and scenarios, while reinforcement learning models often lack interpretability in evaluation values. This study aims to explore whether a low-dimensional, rule-based state variable reinforcement learning model using spatiotemporal data can effectively analyze soccer tactics. We defined state variables and actions based on interviews with a head coach from the Kanto Soccer Division 1 League. All players and scenarios were modeled, reinforcement learning was conducted, and state-action values (Q-values) were calculated. Results revealed a trade-off between action prediction accuracy and TD error. Although no correlation was found with traditional metrics such as xG or goals, the model effectively evaluated high-risk tactics like counterattacks and defensive breakthroughs compared to the baseline. Furthermore, visualizing Q-values and state variables enabled the interpretation of Q-values and evaluation of individual players. This approach shows promise for enhancing coaching strategies and player assessment.