Offline reinforcement learning for badminton tactical decision-making

Abstract

Sports data mining is becoming increasingly vital in modern competitive sports, driven by the need for athletes to continuously enhance their performance. Traditional methods of analyzing sports data rely heavily on expert experience and manual effort, which can be inefficient and unreliable. With advancements in artificial intelligence (AI), sports data is now being processed autonomously, providing more quantitative insights and more comprehensive analysis. This paper focuses on the role of tactics in sports, particularly in badminton, and explores the potential of using AI to enhance badminton tactical decision-making. We investigate the application of offline reinforcement learning (Offline RL) to develop tactical policies from pre-collected datasets, addressing challenges including algorithm design and offline policy evaluation. Specifically, we propose a new variant of conservative Q-learning (CQL), tailored for the hybrid action space to train tactical policies using the integrated offline dataset Shuttle. To evaluate these policies, we develop a preference-based reward model that aligns with tactical preferences, offering an alternative to traditional offline policy evaluation methods. Our computer-based experimental results and analysis demonstrate that the proposed method achieves higher average rewards than all baseline methods and the behavior policy used for data collection. This underscores the potential of the proposed method to enhance badminton tactical decision-making and offer athletes more effective tactical recommendations. Code and data are available at https://github.com/Wenminggong/Offline_RL_for_Badminton. Graphical abstract: [Figure not available: see fulltext.]