Building algorithm portfolios for memetic algorithms
The present study introduces an automated mechanism to build algorithm portfolios for memetic algorithms. The objective is to determine an algorithm set involving combinations of crossover, mutation and local search operators based on their past performance. The past performance is used to cluster algorithm combinations. Top performing combinations are then considered as the members of the set. The set is expected to have algorithm combinations complementing each other with respect to their strengths in a portfolio setting. In other words, each algorithm combination should be good at solving a certain type of problem instances such that this set can be used to solve different problem instances. The set is used together with an online selection strategy. An empirical analysis is performed on the Quadratic Assignment problem to show the advantages of the proposed approach.
