Human-like decision making for lane change based on the cognitive map and hierarchical reinforcement learning

Human-like decision making is crucial to developing an autonomous driving system (ADS) with high acceptance. Inspired by the cognitive map, this paper proposes a hierarchical reinforcement learning (HRL)-based framework with sound biological plausibility named Cog-MP, which combines the cognitive map and motion primitive (MP) in human-like decision making. In the proposed Cog-MP, three general levels involved in ADS are integrated in a top–bottom way, including operational, decision-making, and cognitive levels. The proposed Cog-MP is used to make human-like decisions in lane-changing scenarios, focusing on three aspects: human-like lane decision, human-like path decision, and decision optimization. The proposed framework is validated on two groups of realistic lane-change data, of which one group is used to train cognitions towards different styles of driving behaviors, and the other group is to provide validation scenarios. Experimental results show that the proposed framework can generate human-like decisions and perform soundly regarding the three considered aspects, demonstrating a promising prospect in developing a brain-inspired human-like ADS.