Building a Multi-Agent Deep Reinforcement Learning Framework for Flexible Job Scheduling in Smart Factories 

In this project, we develop an accurate, efficient and scalable model for computing optimal flexible job schedules by using multi-agent reinforcement learning will deep learning techniques for manufacturing factories, in which each machine is considered as an agent. The actions of an agent consist of job routing and job sequencing and its action is decided based on a reward function. In flexible job scheduling, an operation of jobs can be processed by any machine from a given set of alternative machines rather than a fixed machine. The key objectives of this project are (1) designing a factory environment for multi-agent reinforcement learning, (2) building a deep multi-agent reinforcement learning framework for flexible job scheduling, (3) integrating the framework with an enterprise resource planning system with visualization tools, and (4) evaluating the framework in real factory settings. In the framework, each agent learns its own action-reward function independently and simultaneously based on the well-known Q-learning reinforcement learning algorithm with our proposed optimizations, and it is further enhanced through deep learning techniques to improve its accuracy, efficiency and scalability for a large problem size. The developed framework can optimize job schedules for factories to significantly improve their cost effectiveness and productivity that are essential success factors for in the era of Industry 4.0.