Research On Multi-Robot Task Assignment Method Based On Reinforcement Learning

With the rapid development of modern science and technology,there are a lot of problems that need to be solved by multi-robot task allocation technology.The problem of multi-robot task assignment and the multi-agent reinforcement learning algorithm combined with reinforcement learning have gradually become a hot topic of research.At present,many results have been achieved by using reinforcement learning algorithms to solve the problem of multi-robot task allocation,but in practical applications there are still problems such as long training cycles,excessive state-action space,and dimensional disasters.In response to the above problems,this thesis proposes methods for optimizing multi-robot task allocation by studying hierarchical reinforcement learning algorithms and deep reinforcement learning algorithms.This thesis mainly conducts research from the following two aspects:(1)Aiming at the problem that the state-action space is too large when the reinforcement learning deals with the multi-robot task allocation problem,a task allocation algorithm based on Safe Path Combination Option-HRL is proposed.First,the idea of dividing and conquering by hierarchical reinforcement learning is used to transform the global state-action space into a local state-action space,reducing the dimension of the state-action space.Secondly,using the idea of heuristic function,design the generation of the critical path heuristic algorithm Option,Improve the convergence speed of the algorithm.Finally,add a safety mechanism to the action selection strategy to avoid selecting invalid task assignment combinations,ensure the effectiveness of algorithm training,and improve the learning efficiency of the algorithm.(2)Aiming at the problem of dimensional disasters in the multi-robot task allocation method under the complexity of the environment,this thesis proposes a multi-robot task allocation algorithm based on the Heuristically Accelerated Deep Q Network.First,the use of neural networks instead of the Q value in traditional reinforcement learning solves the limitation of the state-action space of reinforcement learning in the high-dimensional space.Secondly,the trajectory pool is introduced into the Deep Q-Network The selection of heuristic actions in the algorithm improves the convergence speed of the algorithm.Finally,the dynamic exploration factor is introduced into the action selection strategy to ensure that the algorithm explores the unknown space in the environment and improves the learning efficiency of the algorithm.Experiments show that the task allocation algorithm based on SPO-HRL alleviates the problem of excessive state-action space in multi-robot task allocation,improves the convergence speed,and optimizes the task allocation results.The HADQN-based task allocation algorithm successfully alleviates the dimensional disaster problem of multi-robot task allocation in complex environments,and at the same time,there is a significant improvement in convergence speed and task allocation results.