Train small, deploy large: Scaling multi-agent reinforcement learning for multi-stage manufacturing lines

Abstract

We present a novel control framework using Multi Agent Reinforcement learning (MARL) that is scalable in the number of workstations in a multi-stage manufacturing line. We show that the dynamics of any production line, regardless of size, can be decoupled into three fundamental expressions. These expressions capture the dynamics of (1) the first workstation, (2) all intermediate workstations, and (3) the last workstation. This decoupling, combined with observation engineering enables training a characteristic 3-workstation, 2-buffer model using MARL methods, which can then generalize to production lines with $w$ workstations with arbitrary cycle times, buffer capacities and reliability models. A numerical study is then conducted to validate the framework.