Deep Reinforcement Learning for Autonomous Driving with Multiple Expert Demonstrations

Deep reinforcement learning (DRL) has emerged as a promising approach to address challenges in autonomous driving. Recently, reinforcement learning from expert demonstrations has gained considerable attention, offering a synergistic integration of imitation learning (IL) and reinforcement learning to enhance model performance. However, in complex driving scenarios, expert decisions may be suboptimal, leading to inefficient training and susceptibility to local optima. In light of this, we propose a novel DRL-based model, termed PME. In the decision-making process, we introduce an enhanced algorithm built upon Proximal Policy Optimization (PPO), which integrates multiple expert policies to mitigate the negative impact of erroneous expert decisions, thereby significantly improving the overall performance of the autonomous driving model. Experimental results, conducted in the CARLA simulator, demonstrate that the proposed PME model surpasses conventional reinforcement learning algorithms across various driving scenarios. The framework consistently achieves higher success rates under diverse conditions, showcasing improved learning stability and superior generalization capabilities.