Tailoring Job Shop Scheduling Problem Instances Through Unified Particle Swarm Optimization

Abstract

Problem instances are paramount when testing the performance of any learning algorithm. For this reason, it is customary to use widespread problems known as benchmark instances. Nonetheless, these are usually generated disregarding heuristics and their nature. For Job Shop Scheduling problem, researchers have created such instances based on random distributions. This idea may bias conclusions about the algorithm under test since a practitioner can only observe performance from a limited perspective, which may not even reflect real-life situations. However, addressing this issue implies tackling the instance generation problem while considering the nature of the solution approach. Hence, in this work, we propose an instance generator based on the Unified Particle Swarm Optimization algorithm, which can tailor instances to different goals. To validate our approach, we include instances generated to different heuristics and instances tailored to a variety of features. In the first case, we seek to favor or hinder one heuristic whereas doing the opposite for the remaining ones. In the second one, we explore instances with specific feature values. Our data reveal that the proposed approach fulfills the expectations and can effectively deal with different kinds of instances. We analyse the nature of the generated instances and their insights, which can be used to further the study about heuristics and problem features.