Research On Multi Agent Transport System Scheduling Based On Deep Reinforcement Learning

As a labor-intensive manufacturing industry and a life-related civilian production industry,the woodworking furniture manufacturing industry has become its future development direction with "personalized customization and flexible production." This paper aims to build a multi-agent system model based on the woodworking material automated transport system in a woodworking furniture intelligent manufacturing factory,and study the end-to-end scheduling strategy for multi-agent transport system.For a start,this paper makes an abstract model of woodworking material automatic transport system.Starting from the actual needs of woodworking material automatic transport tasks,three multi-agent system scenarios were constructed: cooperative navigation,cargo transport,and collision avoidance.Next,a deep reinforcement learning algorithm(ATAC)based on a strategic attention mechanism is proposed to solve the dimensional disaster problem of deep reinforcement learning in the application of multi-agent system.Firstly,the interaction results between the agent and the environment are stored as training samples through the sample pool.Secondly,the actor-critic algorithm is used as the initial framework,and a centralized critic is trained by centralized learning and decentralized execution.The update of network parameters adopts a dual-network partial inheritance mechanism.Thirdly,the strategic attention mechanism is used to focus on important information selectively.Finally,comparative experiments are designed in three scenarios to verify the effectiveness of the ATAC algorithm in solving the problem of transport scheduling in multi-agent system.Then,a deep reinforcement learning algorithm(ADAC)based on an adaptive exploration mechanism is proposed to solve the problem of low exploration degree of deep reinforcement learning in the application of multi-agent system.Firstly,the dual-network partial inheritance mechanism in ATAC is improved.By analyzing the stability of the neural network,the parameters are adaptively inherited to achieve the effect of accelerating training.Secondly,the maximum entropy model is introduced into the objective function of ADAC to make the strategy more random and enhance the exploration ability of the algorithm.Thirdly,the weight of information entropy decreases gradually with the increase of the exploration degree in the action space,making the algorithm's final objective function return to the calculation of cumulative reward value.Finally,comparative experiments are designed in three scenarios to verify the effectiveness of the ADAC algorithm in solving the problem of transport scheduling in multi-agent system.