PIEDQN: An Active-exploration Framework Capable of Prioritizing Experience Replay in Autonomous Driving

There are still many challenges in utilizing DRL to achieve autonomous driving of vehicles in complex environments. Due to random sampling and local optimum, the agent may need to learn efficient navigation strategies. To address these two issues, we propose an active-exploration framework that prioritizes experience replay in autonomous driving. First, we introduce the prioritized experience replay (PER), which ranks the sample experiences in order of importance and improves the policy model's performance. In addition, we introduce the intrinsic curiosity module (ICM) to generate intrinsic reward signals. It enhances the exploration efficiency by encouraging the vehicle to actively interact with the surrounding environment and avoid falling into local optimum. We conducted extensive experiments on the highway-env platform and compared it with other DRL works. The experiment results show that our proposed method improves the average driving speed, the average number of steps, the average cumulative reward, and the average trigger frequency.