Rescheduling Under Disruption: A Product-Driven Framework with Heuristic and Reinforcement Learning Strategies

Abstract

In modern manufacturing, addressing disruptions across multi-stage production requires adaptive and intelligent scheduling. This study evaluates two rescheduling strategies within a product-driven system for the Job Shop Scheduling Problem under disturbances: one based on the Shifting Bottleneck Heuristic (PDS-SBH), and another using a Monte Carlo Reinforcement Learning agent (PDS-RL). Products act as intelligent agents capable of autonomous decisions. A total of 151 simulations were conducted across 14 benchmark instances, with machine-level disruptions modeled as 100%, 200%, and 300% increases in processing times. PDS-SBH achieved average makespan reductions up to 5.2%, serving as a reactive and interpretable baseline. In contrast, PDS-RL consistently outperformed it, achieving reductions of 22.12%, 37.13%, and 53.87%, respectively. These results highlight the superior adaptability of reinforcement learning in uncertain production contexts. The study contributes to the understanding of how combining product-driven architectures with heuristic and learning-based strategies enables the development of intelligent, autonomous, and resilient scheduling systems.