On composing an algorithm portfolio

Evolutionary algorithms are versatile optimization techniques inspired by processes in nature. So far, a wide variety of algorithms have been suggested. However, there is relatively little effort on studying how individual algorithms can work together in a portfolio to achieve a synergy. In this paper, we propose a general methodology to automatically compose a good portfolio from a set of selected evolutionary algorithms. As a single algorithm is a degenerate portfolio, our method also provides an answer to when a portfolio of two or more algorithms are beneficial. Our method has the nice property of being parameter-less; it does not introduce extra parameters. Hence there is no need for parameter control. To illustrate our ideas, we show how a portfolio of five state of the art evolutionary algorithms is automatically constructed using the test functions from the special session on real-parameter optimization of Congress on Evolutionary Computation 2005. It is found that the resulting portfolio obtains the best average ranking. The applicability and limitations of the paradigm of using a benchmarking suite to access evolutionary algorithms are also examined. Though this paper has used evolutionary algorithms only to compose an algorithm portfolio, the idea is generic and is applicable to portfolios with non-evolutionary algorithms as well.