Research On Multi-agent Task Planning Method Based On Prior Knowledge

Investigation is an important multi-agent collaboration application.Compared with one single agent,multi-agent system can greatly improve the search efficiency.The scope and efficiency of the search are much better than that where using one single agent,especially in the situation where the area is large,complex,and contains multiple targets.However,in the existing research on multi-agent collaboration,the number of agents is only a few.When the number of agents is expanded to dozens or even hundreds,the efficiency of the algorithm is greatly reduced.Therefore,how to assign tasks to each agent reasonably to complete the task efficiently is a promising research topic in the field of multi-agent collaboration.For the problem of multi-agent task allocation,there are two main challenges in its research。1)For the task planning of multi-agents,global task planning should consider the task completion efficiency of a single agent;at the same time,the task completion efficiency of an individual agent depends on the assignment of tasks.This two influence each other.2)How to solve the problem of increasing problem solution space as the number of agents increases.In view of the above problems,this paper proposes a bi-level optimization algorithm based on Bayesian optimization and genetic algorithm that incorporates prior knowledge of the probability of occurrence of the target.Firstly,the problem of search task area allocation for large-scale agents.The search area is divided into regular hexagons according to the search radius of the agent.The upper layer algorithm assigns all regular hexagons as tasks to each agent.A genetic algorithm is used in lower layer to mutate the search path of the single agent to obtain the best path.The best path is passed to the upper layer,and the evaluation function of the upper task assignment is revised.And then,in order to speed up the solution of the model,the upper layer uses Bayesian optimization to evaluate the minimum time for eachagent to complete the assigned task,thereby speeding up the convergence of the algorithm.This paper verifies the optimization essence of the algorithm through theoretical derivation,and verifies the collaboration ability of hundreds of agents through simulation experiments.The final experiment shows that,compared with the heuristic algorithm,the algorithm in this paper can shorten the expected path by up to 2.66 times.In addition to theoretical derivation and experimental verification,this paper has designed and implemented a multi-agent task planning visualization tool to formalize the overall search process,which is convenient for determining the convergence trend of the algorithm and displaying the results.