Research On Multi-agent Cooperative Modeling Method Based On Reinforcement Learning

In the background of "Made in China 2025",this paper tends to realize intelligent manufacturing,and improves the level of manufacturing in the uncertain situation of globalization through integrating industrial scenarios and Internet technology.At present,industrial intelligent manufacturing faces common problems such as difficulty in resource sharing,difficulty in process coordination,and dependence on domain experts.Based on agent,multi-agent system and reinforcement learning and other related technologies,a multi-agent model of industrial hybrid intelligent control is constructed,and the agent cooperation ability,flow shop scheduling policy and other key issues in the multi-agent system are studied,in order to realize the industrial production process intelligent control.Based on the multi-agent technology,this paper abstracts the main key points and key entities of the industrial process into individual agents,and establishes a three-tier hybrid multi-agent intelligent model.The model is based on the JADE framework to realize the underlying basic functions.In addition,through some interfaces,the intelligent module of the multi-agent collaboration method based on state representation learning and the intelligent module of the multi-agent system task scheduling method based on reinforcement learning are realized to meet industrial requirements.In this paper,we have improved the M3 DDPG algorithm based on the complex environment that the Agent contacts in the industry,the poor robustness,and the lack of the agent's ability to perceive important features.At the same time,the acquisition of important features by the equipment agent is improved,so that the agent's actions can achieve the expected effect.The improved algorithm uses state representation learning to help capture features,constructs a mapping between observations and state values through deep neural networks,and then the Actor and Critic networks in M3 DDPG learn from the new neural network instead of learning from the initial observations.So that the Agent's actions can reach expectations and can adapt to high-dimensional data.For the flow shop scheduling problem in a multi-agent system,this paper proposes a TS?Qlearning algorithm that combines the tabu search algorithm and Q-learning algorithm.The early training experience of the tabu table storage algorithm of the tabu search algorithm is used to guide the early training of the algorithm.And change the policy of Q-learning algorithm in solving scheduling problems.Based on the exploration advantages of the Q-learning algorithm,our method guides the algorithm in the initial training stage,thereby improving the quality of algorithm training better optimizing resource allocation.