Success-History Based Parameter Adaptation in MOEA/D Algorithm

In this paper two parameter self-adaptation schemes are proposed for the MOEA/D-DE algorithm. These schemes use the fitness improvement ration to change four parameter values for every individual separately, as long as in the MOEA/D framework every individual solves its own scalar optimization problem. The first proposed scheme samples new values and replaces old values with new ones if there is an improvement, while the second one keeps a set of memory cells and updates the parameter values using the weighted sum. The proposed methods are testes on two sets of benchmark problems, namely MOEADDE functions and WFG functions, IGD and HV metrics are calculated. The results comparison is performed with statistical tests. The comparison shows that the proposed parameter adaptation schemes are capable of delivering significant improvements to the performance of the MOEA/D-DE algorithm. Also, it is shown that parameter tuning is better than random sampling of parameter values. The proposed parameter self-adaptation techniques could be used for other multi-objective algorithms, which use MOEA/D framework.