Risk-Augmented Loss With Attention Mechanisms: Enhancing Proximal Policy Optimization for Safer Autonomous Driving

Autonomous driving (AD) on highways presents significant challenges due to dynamic environments, high speeds, and uncertainties arising from sensor noise or abrupt vehicle behavior. This paper explores enhancing decision-making for autonomous vehicles (AVs) using an improved Proximal Policy Optimization (PPO) based on Deep Reinforcement Learning (DRL). The improvement is achieved through a novel risk-augmented loss function. Additionally, attention mechanisms are integrated to enhance performance. Unlike traditional approaches embedding risk in reward functions, our method integrates probabilistic risk assessment directly into the Proximal Policy Optimization (PPO) algorithm’s loss function, ensuring a clear separation between safety and performance objectives. Adaptive and robust decision-making are realized by paying more attention to important driving characteristics and punishing unsafe actions with risk-augmented loss. The experimental results show successes in safety, efficiency, and stability, proving that the proposed approach is effective in realistic highway driving scenarios. This system allows us to develop AV technology and address some of its major hurdles like dynamic environments, reliability and safety.