Evolutionary Learning in Decision Making for Tactical Lane Changing

Decision making for lane change is an important challenge for automated vehicles, especially in complex traffic environments. In recent years, there have been studies that utilize reinforcement learning for lane change applications. However, such an approach requires high computational costs and is difficult to implement by parallel computing. To overcome the problem, an evolutionary learning approach is put forward for the decision-making application of autonomous driving. By deploying the parallel workers, making parameters of the neural network mutate and recombining the well-behaved off-springs during the evolutionary learning process, the Evolution Strategy (ES) agent learns to make decisions for lane-change maneuvers. At the same time, safety verification is performed, which ensures driving safety and simplifies the learning process. To test the performance of the proposed method, a highway simulation environment is established. The results show that the combination of the high-level evolutionary learning and low-level safety verification jointly achieve efficient driving behavior control.