A generic method to compose an algorithm portfolio with a problem set of unknown distribution

As an individual algorithm rarely outperforms all kinds of optimization problems, algorithm portfolios are proposed to combine algorithms and take advantage of their strengths which fits well the prevalent theme of memetic computing. When there are many algorithms to choose from, the possibilities of algorithm combinations are numerous. Therefore, composing an algorithm portfolio which performs well for a given problem class is essential. In this paper, based on a problem set drawn from any unknown problem class according to an unknown probability distribution, we propose a general method to automatically accomplish portfolio construction. The problem set is used as training data for our method to learn an algorithm portfolio suitable for solving the underlying problem class. To construct the portfolio, algorithms are chosen and added one by one. We first find the best-performing algorithm based on its average rank of solving the training problem set. Its most complementary algorithm is then selected by applying the Pearson correlation coefficient of fitness values at the first hitting time. The method iterates to compose the portfolio with more and more algorithms until there is no more improvement. The experimental results indicate the effectiveness of this approach to select well-cooperated algorithms, and the composed portfolio is shown to have the best rank compared to individual algorithms, elite portfolios and comparison algorithms.