Effective and Interpretable Rule Mining for Dynamic Job-Shop Scheduling via Improved Gene Expression Programming with Feature Selection

Abstract

Gene expression programming (GEP) is frequently used to create intelligent dispatching rules for job-shop scheduling. The proper selection of the terminal set is a critical factor for the success of GEP. However, there are various job features and machine features that can be included in the terminal sets to capture the different characteristics of the job-shop state. Moreover, the importance of features in the terminal set varies greatly between scenarios. The irrelevant and redundant features may lead to high computational requirements and increased difficulty in interpreting generated rules. Consequently, a feature selection approach for evolving dispatching rules with improved GEP has been proposed, so as to select the proper terminal set for different dynamic job-shop scenarios. First, the adaptive variable neighborhood search algorithm was embedded into the GEP to obtain a diverse set of good rules for job-shop scenarios. Secondly, based on the fitness of the good rules and the contribution of features to the rules, a weighted voting ranking method was used to select features from the terminal set. The proposed approach was then compared with GEP-based algorithms and benchmark rules in the different job-shop conditions and scheduling objectives. The experimentally obtained results illustrated that the performance of the dispatching rules generated using the improved GEP algorithm after the feature selection process was better than that of both the baseline dispatching rules and the baseline GEP algorithm.