Self-Organizing Collaborative Target Search Of Mobile Multi-Agent Based On Reinforcement Learning

Because of the complexity of search environment,there are many problems in the process of multi-agent search,such as low search efficiency and search repetition.However,the mobile multi-agent system is relatively flexible,which can fully reflect human social intelligence and more adapt to the open and dynamic real environment.Therefore,this thesis studies a kind of multi-agent system based on reinforcement learning to achieve the collaborative search of the target group,so that it can operate normally without human intervention,which is of practical significance to improve the search efficiency of multi-agent collaborative cooperation.This thesis aims at efficient collaborative search of multiple agents,based on the improvement of mobile ad hoc network routing protocols,combined with reinforcement learning methods and special search coverage strategies.Firstly,design a self-organizing network routing protocol based on the improved link state protocol,which is more suitable for multi-agent target search tasks.Then,design a multi-agent cooperation method based on traditional deep deterministic strategy gradient algorithm optimization.Finally,design a mobile multi-agent area coverage search strategy based on reinforcement learning.The innovation work of this article is mainly reflected in the following three parts:(1)Combined with the characteristics of mobile multi-agent self-organizing network,this thesis analyzes the characteristics of multi-agent self-organizing network topology,designs an improved multi-agent self-organizing network routing protocol based on the optimized link state protocol,which is more suitable for the target search scenario.Then,optimizes the multi-agent system based on the protocol and tests the performance of the routing protocol.The test shows that after the improvement the protocol reduces the packet loss rate when the network topology changes,and improves the efficiency of communication mechanism of multi-agent system.(2)By embedding the improved actor critical algorithm into the traditional depth deterministic strategy gradient algorithm and applying it to the iterative steps of multi-agent,a mobile multi-agent collaborative strategy based on the improved depth deterministic policy ladder algorithm is designed to make it more suitable for complex and variable multi-agent target search environments.And use Gym which was provided by open AI to created the experimental environment for testing.The test results show that the improved algorithm can find all kinds of cooperation strategies on the physical and information level in the same environment,and has higher robustness than the existing methods.(3)Bayesian rules are used to design and update the target existence probability map for each agent.Then,a multi-agent region target detection and update algorithm is proposed to fuse the target information observed by each agent with probability map to obtain the exact location of the target group.At the same time,a multi-agent area coverage search algorithm is designed so that the whole multi-agent network can maintain both connectivity and efficiency in the search process.Finally,the simulation is performed on Matlab.The experimental results show that after multiple continuous iterations,the mobile multi-agent system has significantly improved the search of the target group in terms of network stability and area coverage strategy.