Multi-Agent Reinforcement Learning for Schedule-Constrained Automation
Track: Work in Progress
Categories: Transportation, Hospitals / Emergency Services, Supply Chain, Reliability of Critical Systems, Fundamental Research on AI/ML methods for CI
Keywords: multi-agent path finding, multi-agent reinforcement learning, time constraints, schedule execution
Abstract: In modern automation settings, jobs are processed across numerous machines and are characterized by strong inter-task dependencies while adhering to limited equipment availability. When accounting for transportation of jobs be-tween machines, this gives rise to a complex multi-agent routing problem with intricate operational limitations. Existing Multi-Agent Path Finding (MAPF) algorithms used for routing jobs already consider some aspects such as robust-ness and uncertainty but assume instantaneous goal execution. In this paper, we propose MAPF-SC – a lifelong variant of MAPF that in-corporates scheduling constraints for a continuous stream of tasks. We propose to solve MAPF-SC utilizing a Multi-Agent Reinforcement Learning (MARL) formulation with temporal and team reward. We investigate the effects of temporal and topological variations of various automation scenarios on the performance of our method.
Submission Number: 8
Loading