Research On Group Confrontation Strategies Based On Deep Reinforcement Learning

The Multi-Agent Reinforcement learning method has been studied on the basis of game theory and cybernetics before,but the experimental results show that this kind of Multi-Agent Reinforcement learning method can't deal with the complex problems in real life.Until the maturity of deep learning technology and reinforcement learning technology in recent years,new solutions have been brought to the study of swarm intelligence.The deep neural network is used to fit the strategy function,which makes the agent have stronger ability to deal with complex problems.This topic mainly studies the application of multi-agent reinforcement learning method in the environment of confrontation and cooperation,the improvement of algorithm stability and the expansion of the scale of agents,so that agents can cooperate with other agents in the complex environment just like human beings.First of all,the study of this paper is carried out on the GYM platform,which contains many game scenes and is a research platform of supply reinforcement learning method developed by Open AI.In this paper,a swarm intelligence countermeasure strategy is implemented based on MADDPG algorithm.This strategy can interact with the GYM platform,read the information of the agent cluster,and then make a decision to control the agents in the agent cluster to cooperate with each other,so as to achieve the purpose of fighting with other agent cluster.Secondly,this paper improves the DDPG method by changing the deterministic strategy to random strategy to increase the complexity of the samples and improve the stability of the algorithm.It also modifies the strategy gradient update method of DDPG to ensure that the performance of agents remains monotonous in the training process.In this paper,an attention mechanism is added on the basis of the improved MADDPG algorithm,so that a single agent in the intelligent agent cluster can assign different degrees of attention to other agents,so as to reduce the computational complexity of the algorithm and increase the number of intelligent agent clusters.Finally,this paper compares and analyzes the realization difference of MADDPG algorithm,the improved MADDPG algorithm and the MADDPG method after adding attention mechanism on the GYM.The experimental results show that the performance of MADDPG algorithm after adding the attention mechanism is better than that of the traditional MADDPG method.