Pursuit-escape Strategy for Autonomous Agents

The pursuit-escape strategy is the cutting-edge research related to autonomous vehicle. This paper focuses on training 2 agents based on the Q-Learning algorithm model in reinforce learning, in a pursuit-escape game confrontation to solve the game strategy selection problem for a single pursuer and a single escapee in a simulation environment. First of all, the simulation model uses the gym API library provided by the OpenAI research organization to call the pre-defined environment to build the simulation environment. At the same time, the training model of the pursuit-escape game would be built and trained in the cloud computing platform of Baidu PaddlePaddle, and the final training effect that meets the expectation proves that the algorithm is verified. This research also introduces the current interference in to increase the realism of the simulation and optimizes the algorithm related to obstacle judgment. The simulation and experimental results demonstrate that the sensor model can assist in providing "rewards and punishments" in the pursuit process of an agent, which can be applied to improve the autonomous motion strategy selection of the agent.