Collaboration methods for ensembles of dispatching rules for the dynamic unrelated machines environment

Abstract

Dynamic scheduling represents an important combinatorial optimisation problem that often appears in the real world. The difficulty in solving these problems arises from their dynamic nature, which limits the applicability of improvement based metaheuristics. Dynamic problems are usually solved using dispatching rules (DRs), which iteratively construct the schedule. Recently, such heuristics have been constructed using various hyperheuristic methods, most notably genetic programming. Although automatically designed DRs achieve good performance, it is still very difficult to design a single DR that would perform a good decision at every decision point. As a remedy, DRs were combined into ensembles to improve their performance. For that purpose it is required to define how ensembles are constructed and how DRs in the ensemble collaborate. This paper proposes a novel ensemble collaboration method based on a similar method applied for static scheduling problems and adapts it for dynamic problems. The goal is to obtain a collaboration method that produces better results than standard collaboration methods. Additionally, the paper investigates the application of novel ensemble construction methods for dynamic scheduling. The proposed methods are validated on dynamic unrelated machines scheduling problem and compared with existing ensemble construction and collaboration methods. The obtained results demonstrate that the proposed collaboration method performs better than standard ones. Further analyses provide additional insights into the proposed methods and outline several potential research directions in the area of hyper-heuristic ensemble construction.