Uncertainty and Noise Aware Decision Making for Autonomous Vehicles: A Bayesian Approach

Abstract

In the evolving domain of autonomous vehicles, the importance of decision-making cannot be overstated. Deep Reinforcement Learning (DRL) emerges as a pivotal tool in this landscape. However, existing DRL algorithms suffer from inaccurate Q-value estimation, predominantly due to system noise and function approximation errors. This coupled with real-world unpredictabilities, often misdirects autonomous vehicles, leading to sub-optimal actions and safety hazards. This work introduces a novel DRL algorithm tailored for uncertainty and noise-aware decision-making in autonomous vehicles. Our approach harnesses Bayesian Neural Networks and skew-geometric Jensen-Shannon divergence, to rectify the aforementioned limitations. Evaluated on the OpenAI gymnasium environment, our algorithm has clear advantages over existing methods in terms of cumulative rewards and convergence speed.