Safe Reinforcement Learning Strategies with Interpretable Decision-Making for Autonomous Driving in Uncertain Traffic Conditions

The study focuses on improving the safety and interpretability of reinforcement learning in autonomous driving under uncertain traffic conditions. A decision-making model is developed using the Soft Actor-Critic algorithm, with an added module to estimate uncertainty and detect risky situations in real time. To make the system’s behavior more understandable, a state–action salience mapping is designed to show which inputs have the greatest effect on each decision. The model is tested in simulation environments involving sudden pedestrian crossings, lane changes by other vehicles, and complex traffic flows. Results show that the method reduces the accident rate by 23.5% compared with standard approaches, while also making it easier for users to follow the reasoning behind the system’s actions. These findings suggest that combining risk detection with simple visual explanation tools can help reinforcement learning models perform more reliably and transparently in real-world traffic.