Decision-making for Connected and Automated Vehicles in Chanllenging Traffic Conditions Using Imitation and Deep Reinforcement Learning

Decision-making is the "brain" of connected and automated vehicles (CAVs) and is vitally critical to the safety of CAVs. The most of driving data used to train the decision-making algorithms is collected in general traffic conditions. Existing decision-making methods are difficult to guarantee safety in challenging traffic conditions, namely severe congestion and accident ahead. In this context, a semi-supervised decision-making algorithm is proposed to improve the safety of CAVs in challenging traffic conditions. To be specific, we proposed the expert-generative adversarial imitation learning (E-GAIL) that integrates imitation learning and deep reinforcement learning. The proposed E-GAIL is deployed in roadside unit (RSU). In the first stage, the decision-making knowledge of the expert is imitated using the real-world data collected in general traffic conditions. In the second stage, the generator of E-GAIL is further reinforced and achieves self-learn decision-making in the simulator with challenging traffic conditions. The E-GAIL is tested in general and challenging traffic conditions. By comparing the evaluation metrics of time to collision (TTC), deceleration to avoid a crash (DRAC), space gap (SGAP) and time gap (TGAP), the E-GAIL greatly outperforms the state-of-the-art decision-making algorithms. Experimental results show that the E-GAIL not only make-decision for CAVs in general traffic conditions but also successfully enhances the safety of CAVs in challenging traffic conditions.