Multi Objective Learning Classifier Systems Based Hyperheuristics for Modularised Fleet Mix Problem

摘要

This paper presents an offline multi-objective hyperheuristic for the Modularised Fleet Mix Problem (MFMP) using Learning Classifier Systems (LCS). The LCS based hyperheuristic is built from multi-objective low-level heuristics that are derived from an existing MFMP solver. While the low-level heuristics use multi-objective evolutionary algorithms to search non-dominated solutions, the LCS based hyperheuristic applies the non-dominance concept at the primitive heuristic level. Two LCS, namely the extended Classifier System (XCS) and the sUpervised Classifier System (UCS) are augmented by multi-objective reward and accuracy functions, respectively. The results show that UCS performs better than XCS: the hyperheuristic learned by the UCS is able to select low-level heuristics which create MFMP solutions that, in terms of a distance-based convergence metric, are closer to the derived global Pareto curves on a large set of MFMP test scenarios than the solutions created by heuristics that are selected by the XCS hyperheuristic.