Dynamic dispatching rule selection for the job shop scheduling problem

Uncertainty is pervasive in modern manufacturing settings. In order to cope with unexpected events, scheduling decisions are commonly taken resorting to dispatching rules, which are reactive in nature. However, rule performance varies according to shop utilisation and due date allowance, which often change in dynamic real-world job shops. Therefore, this paper explores systems that select dispatching rules as conditions change over time, namely periodic and real-time dispatching rule selection systems, which are based on supervised learning and reinforcement learning algorithms, respectively. These types of systems have been proposed in the past but have been further improved in this work by carefully selecting the most relevant state features and dispatching rules. Moreover, by testing both approaches on the same instances, it was possible to compare them and determine the most advantageous one. After the tests, which included a wide array of job shop instances, both periodic and real-time systems outperformed state-of-the-art dispatching rules by over 10% tardiness-wise. Nonetheless, the periodic rule selection approach was more robust across all tests than the real-time approach. These results demonstrate that there is a real incentive for managers to adopt dispatching rule selection systems. • Systems that dynamically select dispatching rules are explored and enhanced. • Supervised and reinforcement learning based systems are compared. • Superior performance is achieved due to improved state and action sets. • Dispatching rule selection systems outperformed the best rule by over 10%. • Periodic rule selection is more robust than real-time rule selection.