A Reinforcement Learning Method in Cooperative Multi-Agent System for Production Control System

摘要

Nowadays, in various domains including distributed control, telecommunication, robotics and economics to address the problems, Multi-Agent Systems (MAS) is used. To solve with preprogrammed agent behavior complexity of many tasks of these domains making it difficult. The multi- agent systems can limit problem of control in complex system production, to solve more efficiently. However, local optimization tendencies are often shown in iterative development algorithms. So, by using reinforcement learning and considering indicators of global key performance this paper presents a new method for cooperative multi-agent system. To cooperative order agents, for the purpose of this a central deep and double deep learning module its knowledge is being transferred. From the results, subsequent reinforcement learning using memory, the order's experience is stored and increases 50% of average mean cycle. In comparison of production control conventional methods, nonlinear production system and parameters of learning interdependencies between the characteristics are investigated and the performance is evaluated.