DDPG-LSTM Framework for Personalized Athlete Training Plan Optimization and Competition Strategy Generation

Abstract

Traditional methods of athlete training plan formulation often rely on the coach's experience and expert judgment, leading to challenges in dynamically adjusting training plans to athletes' real-time performance and physical conditions. Such static approaches can cause issues like overtraining or undertraining, affecting athletes' overall performance. This paper introduces a deep reinforcement learning (DRL) framework, leveraging real-time data analysis to optimize personalized training plans and automatically generate intelligent competition strategies. By utilizing the Deep Deterministic Policy Gradient (DDPG) algorithm within the Actor-Critic framework, the study employs a state-of-the-art implementation with hyperparameters such as a learning rate of 0.001, batch size of 64, and discount factor (gamma) of 0.99. The key action spaces are defined, including training load (intensity), frequency, and rest intervals, while the reward function is tailored to balance training stress and performance improvement. Additionally, a Long Short-Term Memory (LSTM) model is integrated to analyze time-series data, refining strategies based on dynamic performance feedback. Experimental results show that the DDPG-based approach significantly improves athletes' performance by 12% in key metrics, such as shooting accuracy, and maintains the athletes' Training Stress Balance (TSB) in a healthy positive range over a 90-day training cycle. The LSTM-based game strategies, tested in simulated basketball playoff scenarios, outperform traditional strategies, increasing the final score by 13 points (104 vs. 91), demonstrating substantial improvements in competitive performance and strategy optimization.