Robust Human-Centered Assembly Line Scheduling with Reinforcement Learning

This study set out to develop a Reinforcement Learning (RL) agent for solving an extended Permutation Flow Shop Scheduling Problem (PFSSP). From the domain perspective, we see a lack of realistic constraints for synchronized, human-centered assembly lines. Moreover, objective functions must be provided to enable stress-reducing as well as robust planning under uncertainty. From a methodical perspective, RL has received more and more attention for problems of this type. However, we cannot identify applicable RL concepts for our extended PFSSP with multicriteria objectives. We propose a generic RL agent, which operates on an abstract representation of the schedule and with an objective-independent reward function. Our numerical experiments demonstrate that the agent successfully generalizes a policy and achieves better scores than a Simulated Annealing (SA) metaheuristic.