Research On Multi-agent Roundup Strategy Based On Reinforcement Learning

With the advent of the Internet era,the development trend of artificial intelligence is gradually shifting from single agent decision-making to multi-agent decision-making.There are many multi-agent systems in life,such as driving cars through traffic,playing football,and even bees building hives.These systems are multi-agents through collaboration or competition to achieve a certain goal.Compared with single-agent learning,multi-agent learning challenges lie in its larger search space,perception to other agents and robustness of the system.In multi-agent systems,with the increase in the number of agents,the problem is that it is difficult for agents to distinguish the information conducive to cooperation from the globally Shared information,because the information that does not facilitate cooperation may affect the teamwork.If all agent information is Shared,the action space is too large.At the same time,receiving a large amount of information requires high bandwidth,long delay and high computational complexity.Moreover,when an agent outputs actions according to the observed information and overall strategy,it will produce ambiguity.So finding a way for agents to select information useful for collaboration when needed is particularly important.Aiming at the problems existing in multi-agent reinforcement learning mentioned above,this topic constructs a multi-agent roundup environment to study multi-agent reinforcement learning,and the main research contents are as follows:(1)This paper summarizes the research status of multi-agent reinforcement learning at home and abroad,classifies algorithms from the perspective of single agent and multi-agent,introduces the classical algorithm and model structure,and summarizes and forecasts the application prospect and development trend of multi-agent reinforcement learning cooperation or competition technology.(2)In order to solve the problem that multi-agent system is difficult to extract information with the expansion of joint information space,a multi-agent reinforcement learning strategy based on filtering mechanism filtering information(FMAC)is proposed.Firstly,the agent information is coded,and then the irrelevant agent information is filtered by calculating the degree of association between agents,so as to realize effective communication between agents in cooperative environment.In addition,the method of centralized training and decentralized execution is adopted to solve the non-stationarity of environment.To contrast algorithm proves the advantage of the improved algorithm to the multiagent Open AI particle envs as test platform,in which the simple spread were compared with simple tag environment experiment and found that the improved algorithm improved strategy iterative efficiency and generalization ability,in the case of agent populations remained stable effect,helps to multi-agent reinforcement learning is applied to a wider range of areas.(3)In order to improve the ability of cooperative strategy of agents,this paper set a multi-agent roundup single agent experiment environment,which was improved on the basis of Simple Tag to improve the complexity of the environment.This paper adopts the idea of the training of the encirclement network and the escape network to improve the difficulty and tactics of the encirclement by improving the ability of the escapers.The experiment proves that the countermeasure training can improve the strategy of rounding up and the generalization ability.