Comparing Selection Hyper-Heuristics for Many-Objective Numerical Optimization

Abstract

Mechanisms for automatic selection of parameters/heuristics used by evolutionary algorithms can provide more robust and independent approaches. In this work we propose an approach composed of a selection hyper-heuristic implemented within the MOEA/DD (Multi-objective Evolutionary Algorithm based on Dominance and Decomposition) algorithm based on Differential Evolution. Four selection hyper-heuristics are considered in this study: Thompson Sampling, Probability Matching, Adaptive Pursuit and Self-Adaptive Differential Evolution. The hyper-heuristics are employed to choose the crossover operator selected from a pool of operators, according to a probability that reflects the operator's previous performance during the evolutionary process. The MaF benchmark is considered with 5, 10 and 15 objectives. This benchmark includes a diversity of characteristics, representing the challenges that real-world problems may pose. Statistical tests indicate that the proposed approach performs equally or even outperforms those with fixed crossover operator.