Does size matter? On the influence of ensemble size on constructing ensembles of dispatching rules

Abstract

Recent years saw an increase in the application of genetic programming (GP) as a hyper-heuristic, i.e., a method used to generate heuristics for solving various combinatorial optimisation problems. One of its widest application is in scheduling to automatically design constructive heuristics called dispatching rules (DRs). DRs are crucial for solving dynamic scheduling environments, in which the conditions change over time. Although automatically designed DRs achieve good results, their performance is limited as a single DR cannot always perform well. Therefore, various methods were used to improve their performance, among which ensemble learning represents one of the most promising directions. Using ensembles introduces several new parameters, such as the ensemble construction method, ensemble collaboration method, and ensemble size. This study investigates the possibility to remove the ensemble size parameter when constructing ensembles. Therefore, the simple ensemble combination method is adapted to randomly select the size of the ensemble it generates, rather than using a fixed ensemble size. Experimental results demonstrate that not using a fixed ensemble size does not result in a worse performance, and that the best ensembles are of smaller sizes. This shows that the ensemble size can be eliminated without a significant influence on the performance.