Comparison of ensemble learning methods for creating ensembles of dispatching rules for the unrelated machines environment

Dispatching rules are often the method of choice for solving various scheduling problems, especially since they are applicable in dynamic scheduling environments. Unfortunately, dispatching rules are hard to design and are also unable to deliver results which are of equal quality as results achieved by different metaheuristic methods. As a consequence, genetic programming is commonly used in order to automatically design dispatching rules. Furthermore, a great amount of research with different genetic programming methods is done to increase the performance of the generated dispatching rules. In order to additionally improve the effectiveness of the evolved dispatching rules, in this paper the use of several different ensemble learning algorithms is proposed to create ensembles of dispatching rules for the dynamic scheduling problem in the unrelated machines environment. Four different ensemble learning approaches will be considered, which will be used in order to create ensembles of dispatching rules: simple ensemble combination (proposed in this paper), BagGP, BoostGP and cooperative coevolution. Additionally, the effectiveness of these algorithms is analysed based on some ensemble learning parameters. Finally, an additional search method, which finds the optimal combinations of dispatching rules to form the ensembles, is proposed and applied. The obtained results show that by using the aforementioned ensemble learning approaches it is possible to significantly increase the performance of the generated dispatching rules.