Hard computational problems are often solvable by multiple algorithms that each perform well on different problem instances. We describe techniques for building an algorithm portfolio that can outperform its constituent algorithms, just as the aggregate classifiers learned by boosting outperformthe classifiers of which they are composed. We also provide a method for generating test distributions to focus future algorithm design work on problems that are hard for an existing portfolio. We demonstrate the effectiveness of our techniques on the combinatorial auction winner determination problem, showing that our portfolio outperforms the state-of-the-art algorithm by a factor of three.

PDF file

Details