Interpretable deep reinforcement learning for hybrid flowshop scheduling under sequence dependent family setup time, unrelated machines and transportation time constraints

Abstract

Modern manufacturing environments require more than static production schedules: they demand adaptive and dynamic solutions capable of responding to frequent, irregular job orders and unforeseen events on the shop floor. To meet these challenges, schedules must be regenerated both rapidly and efficiently as new demands arise. This underscores the need for automated optimisation methods that can learn from experience, adapt to changing conditions, and update solutions at a pace that matches or exceeds the rate of change within the manufacturing system. This thesis explores the potentials of MatNet, a recent deep reinforcement learning (DRL) model, applied as a black box scheduler for generating high-quality schedules in realtime. While MatNet has shown promise in various combinatorial optimisation problems, its internal decision-making processes remain largely unexplored. This research aims to bridge that gap by systematically evaluating MatNet’s performance within an industrial scheduling context. A highly constrained hybrid flowshop scheduling problem with sequence dependent family setup times and transportation costs serves as a practical test case. To improve interpretability, Sobol sensitivity analysis is also employed to evaluate and provide insight into MatNets behaviour as a black-box scheduler. The results show that the MatNet-based DRL approach can effectively learn robust scheduling strategies, scale to larger problem instances beyond those seen during training and generate high-quality solutions within seconds. These capabilities demonstrate its suitability for deployment in real-world manufacturing environments where rapid response to unforeseen changes is critical