Evaluating Player Performance and Tactical Decision-Making in Racket Sports Using Deep Reinforcement Learning

This paper introduces a novel evaluation network that leverages deep reinforcement learning integrated with advanced modeling and prediction strategies to enhance decisionmaking in racket sports, drawing parallels to challenges encountered in unmanned autonomous systems (UAS). Traditional performance analysis methods predominantly depend on manual observation and static metrics, which inadequately capture the dynamic and strategic complexities of game environments. Our approach combines the principles of Markov Decision Processes with the Transformer architecture to manage long sequential tasks, thereby improving the accuracy of correlating states and actions. By modeling turn-based racket sports, the evaluation network assigns $Q$-values to actions derived from historical match data, demonstrating the impact of each action on potential scoring or loss of points. We evaluated our approach using various hyper-parameters and network architectures, validating it against multiple baselines and performance metrics. The findings suggest promising avenues for advancing data-driven training methodologies and enhancing autonomous system capabilities in complex, task-oriented domains.