摘要：AbstractThis paper proposes a distributed reinforcement learning method based on deep Q-network and the consensus algorithm to deal with the multi-vehicle platoon control problem, which contains the two processes of local training and global consensus. The platooning problem is decomposed into many single-vehicle tasks based on deep Q-network, where each vehicle accumulates its experience data samples by interacting with its front and back vehicles. After initialization, all vehicles’ Q-networks are first locally optimized based on their own experience simultaneously. The consensus algorithm is then used to make all vehicles in a decentralized platoon approach each other, where the communication is only required among directly connected vehicles. At last, the simulation study shows that the Q-networks of all vehicles reach consensus first and then converge to the optimum in union using the proposed distributed deep Q-networks algorithm, and all vehicles learn to form the required platoon and move forward with a roughly equal separation.
