Adaptive training load optimization for track and field athletes: A reinforcement learning approach

Optimization of a training program for an athlete is a difficult issue in sport science. It is a delicate trade-off between stimulating performance enhancement and allowing adequate recovery to prevent injury. The paper, presents a detailed structure of the offline optimization of training loads using the DQN architecture. The framework overcomes the simulation gap by applying a data-driven transition model as a digital twin, which can be used to find the optimal training policies without the ethical or safety concerns of conducting the experiment in real-time on athletes. This intelligent model, founded on comprehensive physiological and performance data collected from 25 athletes over a whole training season, has the capacity to dynamically provide ideal training prescriptions like increased intensity, increased volume, or active recovery. Considered data include parameters such as Heart Rate Variability (HRV), sleep quality, training loads, Acute to Chronic Workload Ratio (ACWR), and weekly performance. The proposed architecture uses a feedforward neural network as an estimator of the Q-value function. By optimizing the adaptive -Greedy policy and the Experience Replay Buffer, the stability and efficiency of the learning process are ensured. In addition, a dual reward function including performance reward and physiological state reward is designed. This function guides the agent towards policies that simultaneously lead to short-term performance improvement and long-term health maintenance of the athlete. The experimental results show that the model has successfully reduced the error rate and has tended to converge to near zero for the loss function. Also, the proposed method has shown a high ability in managing training load and controlling the risk of injury, in such a way that it has been able to dynamically and with high adaptability reduce the risk and always maintain the performance of athletes within the optimal range.