Multi-agent-based deep reinforcement learning for dynamic flexible job shop scheduling

In disruption-prone manufacturing environments, flexible job shop scheduling becomes a dynamic problem. For achieving a high solution quality, operations research approaches can be applied. In contrast, due to the required fast response times, dispatching rules are the standard. In order to elaborate on both, we present a new deep reinforcement learning algorithm. It combines policy gradient algorithms with actor-critic architectures and interprets the production system as a multi-agent system. Our evaluation on benchmark instances shows that the algorithm generates better schedules than dispatching rules with the same response time. It also generalizes well when tested on different manufacturing environments.