Effective dispatching rules mining based on near-optimal schedules in intelligent job shop environment

In many intelligent job shop, a scheduling system has collected and stored huge amounts of production data. Meanwhile, managers still believe that more scheduling knowledge should be further acquired from historical production data. To this end, this paper focuses on exploring newly dispatching rules via near-optimal schedules generation, data transformation, and dispatching rules mining. Without loss of the optimality, we propose a swap operation-based method to improve the quality of feasible schedules, generated by existing dispatching rules or meta-heuristic algorithms. Then, data transformation converts them into training data set with four types of constructed attributes. Thirdly, we present a hybrid genetic algorithm and random forest to extract high-quality training data set and evolve dispatching rules. The experimental results indicate that the proposed approach is effective and robust, and the newly dispatching rules are superior to the existing dispatching rules and some learning algorithms. Finally, a case study with two scenarios in a real environment also shows that the proposed approach outperforms classical rules and meta-heuristic algorithms.