多智能体动态目标协同搜索策略研究

Research on multi⁃agent collaborative search strategy of escaping target

针对目前多智能体集群搜索采用的分区域巡逻策略在搜索具备躲避能力的动态目标时,在分区边界地带搜索效果不佳、巡逻规律易被掌握、无法应对单体故障导致的其管辖区域成为盲区等问题,提出了一种运用强化学习框架的协同搜索策略。该方法通过对作为搜索者的多智能体和随机生成并具有躲避策略的目标进行对抗训练,最终训练出能指导智能体行为的协同搜索策略,优化多智能体集群搜索系统的搜索表现。在三维仿真平台Gazebo中对最终训练得到的协同搜索策略进行仿真,结果表明集群搜索系统应用该协同搜索策略比应用分区域巡逻策略有更高的搜索效率、更高的随机性和更强的鲁棒性。

A collaborative search strategy based on the reinforcing learning framework is proposed to improve the performance of the current strategy in which agents patrol their respective districts to find escaping target.There are many problems of current strategy including the poor performance to search the boundaries between districts,the patrol patterns are easy to acquire,and the lack of response from the blind zone caused by agents’malfunction.To solve the above problems,adversarial training between agents as seekers and target with an escape strategy is executed to produce a collaborative search strategy which guides agents'actions.With the implementation of the strategy after training in Gazebo three-dimensional physical simulation platform,new collaborative search strategy is verified to have better overall performance,higher complexity,and stronger robustness com‑pared with the current patrol strategy.