Dynamic Preference for Autonomous Driving: A Deep Reinforcement Learning Approach

Deep reinforcement learning (DRL) based autonomous driving vehicles (ADVs) has attracted numerous researchers in recent years. Many DRL‐based urban traffic scenarios ADVs have been studied. However, there still remains a notable gap in the literature regarding the incorporation of dynamic passenger preferences. For instance, the passenger dynamic preferences between safety, comfort and efficiency change as traffic changes. Existing researches often overlook the dynamic nature of passenger preferences, which can significantly impact the overall passenger experience. Therefore, to improve driving personality and efficiency, in this paper, we deal with the challenge of incorporating dynamic passenger preferences into ADVs using DRL. We propose an innovative DRL method that first leverages prior knowledge to reduce the preference space which can speed up the training process. Then we apply homotopy optimization to make DRL optimization problem gradually transition from non‐preference to dynamic preference. We validate our approach utilizing a simulated merge environment. Experimental results show that our approach has faster convergence and better performance when dealing with dynamic preference.