Robust Decision Making for Autonomous Vehicles at Lane-Merging Scenarios: A Multi-Objective Deep Reinforcement Learning Approach

Deep reinforcement learning (DRL) has demonstrated advanced ability to deal with complex decision making problems in autonomous driving. However, the driving demands (preferences) of passengers are hardly considered. In addition, the preferences usually change over time, how to balance these preferences still remain challenge for traditional DRL. In this paper, we present a multi-objective DRL approach considering dynamic preferences. Specifically, we model the sequential decision making problem with dynamic preferences for lane merging as markov decision process. Additionally, the random sampling with prior knowledge is advanced to reduce the sampling space and promote convergence rate. Finally, the homotopy optimization is introduced to avoid overfitting and balance the multiple preferences. We test the effectiveness of our algorithm in a lane-merging scenario. The results show that our method out performs the baseline method.