Abstract: To solve the problem that agents cannot make effective decisions under local observation, a conflict resolution method combined with deep reinforcement learning is proposed. Based on DDQN algorithm, this method uses the characteristics of reinforcement learning mode to calculate the cumulative return of agent and determine the priority of agent through the return value, so as to achieve the purpose of conflict resolution. The method is evaluated by simulating the traffic jam in real life, and the experimental results show that the method can effectively solve the agent conflict.