Application of deep reinforcement learning in sports competitive decision-making

Demand for tactical optimisation and decision support in sports competitions is growing; traditional rule-based methods suffer from poor adaptability and latency.This study builds a DRL-based decision model for Wushu Sanda, trained on multi-source data and validated in simulation.The agent learns policies via interaction to optimise tactical choices in dynamic contexts.Compared with rule-based and classical RL baselines (Q-learning, SARSA), our model achieves higher decision accuracy, larger cumulative reward, and faster convergence.It adapts to diverse scenarios and supports real-time tactical adjustment.We also identify challenges in data quality, computational cost, and cross-sport generalisation.The findings highlight DRL's practicality for competitive decision-making and outline directions for improving interpretability, sample efficiency, and deployment in live matches.