| Abstract | Badminton performance evaluation has traditionally relied on subjective coaching assessments and basic statistical metrics, limiting scalability and real-time feedback. This study proposes a novel framework leveraging Proximal Policy Optimization
(PPO)—a reinforcement learning (RL) algorithm—to automate and enhance player performance analysis through data-driven decisionmaking. By integrating multi-modal inputs (computer vision for shuttle tracking, feedback from players and match statistics), the system trains PPO agents to evaluate tactical choices (e.g., shot selection, footwork efficiency) and strategic adaptability during rallies [1]. The PPO-based model dynamically optimizes a reward function that quantifies player strengths and weaknesses, balancing
short-term actions (e.g., smash effectiveness) with long-term game outcomes (e.g., rally win probability). Key performance metrics include a Player Skill Score (PSS)—a composite AI-generated rating—and policy convergence speed, demonstrating PPO’s superiority in stability and adaptability. Additionally, the system enables opponent modeling and personalized training recommendations by simulating adversarial strategies. Results show that the PPO-driven system outperforms traditional evaluation methods in accuracy and granularity, as verified by coach assessments. This work bridges sports science and AI,
offering a scalable, objective tool for badminton performance optimization, with potential extensions to other racket sports.
Keywords: Reinforcement Learning, PPO, Badminton Analytics,
Sports AI, Multi-Modal Data Fusion
|
|---|
