Counterfactual Evolutionary Reasoning for Virtual Driver Reinforcement Learning in Safe Driving

Safety is the primary concern in the motion planning and decision-making of the virtual driver that provides prescriptions to the real human driver and even performs self-driving in the absence of human take-over. For such an issue, traditional reinforcement learning methods, limited by their learning mechanisms, suffer from a slow convergence of model training as well as a less consideration for early warning of possible accidents. To address the above deficiency, this paper proposes a new method based on counterfactual evolutionary reasoning that can be used to build the virtual driver. The method treats safe driving as a sequential decision-making problem with sparse rewards, and employs counterfactual evolutionary reasoning to guide the searching direction as well as to accelerate the model training. An intervention mechanism from outlier distributions is further introduced to enhance the model's ability of exploration. Experiments in the virtual test environment indicate that the proposed method, compared with other typical reinforcement learning techniques, both achieves a higher safe arrival rate and a faster convergence speed.