Flexible Job-Shop Scheduling Algorithm And System Based On Multi-Agent Deep Reinforcement Learning

With developments in economic globalization and information technology,minibatch-size,customize manufacturing process becomes a major requirement at the high end of the market,where scheduling depends on various orders instead of a few standardized and fixed products.The highly complexed flexible job-shop scheduling problem embarrassed traditional approaches,for instance,mathematical methods are lack of scalability,heuristic methods are costly,and even human planners sometimes fall into disorder.Therefore,this work presents a flexible job-shop scheduling system with solution based on multi-agent deep reinforcement learning.1.In this dissertation,a flexible job-shop scheduling progress is modeled as MultiAgent Markov Decision Progress,where each machine corresponds to an autonomous agent,and jobs in the problem make up the action sets for machine agents.A simulation environment for the flexible job-shop scheduling problem is implemented to provide static information and dynamic responds to agents.Therefore,it constructs an algorithm framework for flexible job-shop scheduling problem by interaction among reinforcement learning agents and simulation environment.2.Three kinds of algorithms based on multi-agent deep reinforcement learning are introduced to learn to solve the task,including Independent Learning,Mean Field Q and Graph policy.Independent Learning algorithm framework introduces two types of single agents deep reinforcement learning methods to the problem.MultiType Mean Field Q algorithm and Graph policy algorithm is devoted to the complex relationship among machines and among jobs respectively.Experiments shows that the proposed methods are of great effectiveness.3.Based on the proposed solution for the flexible job-shop scheduling problem,a software system was proposed,where manufacturing customers could customize their problem and start solving progress easily by a Web App.In addition,the module of solving progress can be apart from the main system to provide research platform for researchers.