Reward design and hyperparameter tuning for generalizable deep reinforcement learning agents in autonomous racing

Deep Reinforcement Learning (DRL) is transforming autonomous racing by enabling agents to make real-time, high-stakes decisions with the least supervision. Yet, strong generalization over multiple varied tracks is a key bottleneck. Within this paper, a rigorous examination of the relationship between reward system design and hyperparameter tuning for autonomous racing agents using the AWS DeepRacer platform as a unified benchmark is conducted. A comprehensive comparison of Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) algorithms on two vastly different reward structures with the aid of an extensive tuning of batch size, learning rate, discount factor, and entropy is performed. The results identify that a well-engineered reward mechanism, under the optimized hyperparameter (batch size 128, learning rate 0.0003, discount factor 0.99, entropy 0.01), allows PPO to outperform standard benchmarks with an average lap time of 12.464 s on 21 unseen tracks. These results demonstrate not only enhanced performance but also improved generalization, enabling the models to perform effectively on previously unseen tracks. Additionally, significant emphasis was placed on reward shaping and analyzing hyperparameter sensitivity in large-scale DRL systems to ensure their practical applicability in autonomous scenarios.