Multiobjective Evolutionary Algorithm Portfolio: Choosing Suitable Algorithm for Multiobjective Optimization Problem

The concept of algorithm portfolio has a long history. Recently this concept draws increasing attention from researchers, though most of the researches have concentrated on single objective optimization problems. This paper is intended to solve multiobjective optimization problems by proposing a multiple evolutionary algorithm portfolio. Differing from previous approaches, each component algorithm in our portfolio method has an independent population and the component algorithms do not communicate in any way with each other. Another difference is that our algorithm introduces no control parameters. This parameter-less characteristic is desirable as each additional parameter requires independent parameter tuning or control. A novel score calculation method, based on predicted performance, is used to assess the contributions of component algorithms during the optimization process. Such information is used by an algorithm selector which decides, for each generation, which algorithm to use. Experimental results show that our portfolio method outperforms individual algorithms in the portfolio. Moreover, it outperforms the AMALGAM method.