Using a Genetic Algorithm-Based Hyper-Heuristic to Tune MOEA/D for a Set of Benchmark Test Problems

摘要

The multi-objective evolutionary algorithm based on decomposition (MOEA/D) is one of the most popular algorithms in the EMO community. In the last decade, the high performance of MOEA/D has been reported in many studies. In general, MOEA/D needs a different implementation for a different type of problems with respect to its components such as a scalarizing function, a neighborhood structure, a normalization mechanism and genetic operators. For MOEA/D users who do not have the in-depth knowledge about the algorithm, it is not easy to implement an appropriate algorithm that is suitable for their problems at hand. In our previous studies, we have suggested an offline genetic algorithm-based hyper-heuristic method to tune MOEA/D for a single problem. However, in real-world situations, users may want to use an algorithm with robust performance over many problems. In this paper, we improve the offline genetic algorithm-based hyper-heuristic method for tuning a set of problems. The offline hyper-heuristic procedure is applied to 26 benchmark test problems. The obtained MOEA/D implementations are compared with six decomposition-based EMO algorithms. The experimental results show that the tuned MOEA/D outperformed the compared algorithms on many test problems. The tuned MOEA/D also shows good (and stable) performance over a set of test problems.