Multi-agent Evasion Algorithm Design Based On Reinforcement Learning

With the improvement of the level of intelligence in modern society,the research on multi-agent has become a hot topic at present.Pursuit-evasion problem of multiagent is one of the core issues in multi-agent research because of both cooperation and competition games.Since the pursuit-evasion game was put forward,the pursuit-evasion problem has gradually developed into a large family of problems.This paper mainly studies the multi-agent evasion problem in three-dimensional environment.In this paper,the method of reinforcement learning is introduced to make up for the deficiency that the traditional method can't design the controller without model.In this paper,a multi-agent evasion algorithm based on DQN algorithm is proposed,which is distributed learning.Through two stages of self-learning,the agent optimizes its own evasion strategy by setting different subtasks.In order to solve the problem that the agent belong to the same camp may occur collision in the multiagent escape algorithm,a multi-agent evasion and collision avoidance algorithm is proposed.At the same time,in order to make the multi-agent evasion and collision avoidance algorithm have universality and generalization ability to any evader,the selection rules of the initial state of multi-agent are designed.The improved algorithm is verified by simulation.In addition,this paper also proposes a centralized multi-agent evasion algorithm to solve the defect of environment instability caused by other agents as part of the environment in distributed learning,which affects the convergence of the algorithm.The improved algorithm is verified by simulation.Based on the analysis of the simulation results,after the training,the multi-agent evasion algorithm converges successfully,and the agent under the pursue can use this evasion strategy to successfully evade in the three-dimensional environment.