Research on intelligent vehicle decision system based on reinforcement learning

Abstract This study presents a decision-making model for autonomous vehicles that integrates deep reinforcement learning (DRL) with a risk correction mechanism. This integration aims to enhance dynamic environmental perception and risk assessment in complex traffic scenarios. Initially, the model constructs driving information for the ego vehicle and surrounding vehicles. Subsequently, it introduces a polarized self-attention (PSA) and bidirectional long short-term memory (BiLSTM) combined safety mechanism (PSA-BiLSTM) to improve attention to potential hazardous vehicles in complex high-speed scenarios. Furthermore, a risk correction module has been designed to evaluate and adjust decision-making actions, ensuring driving safety. The model leverages real-time driving information of the target vehicle and its surroundings to generate vehicle decision commands at each sampling point using DRL. Simulation results demonstrate that this approach significantly improves safety and learning speed compared to pure reinforcement learning methods. Additionally, it exhibits higher driving efficiency and lower accident rates compared to imitation learning and traditional rule-based decision models in complex high-speed scenarios.