Statistical Comparison of Algorithm Performance Through Instance Selection

Empirical performance evaluations, in competitions and scientific publications, play a major role in improving the state of the art in solving many automated reasoning problems, including SAT, CSP and Bayesian network structure learning (BNSL). To empirically demonstrate the merit of a new solver usually requires extensive experiments, with computational costs of CPU years. This not only makes it difficult for researchers with limited access to computational resources to test their ideas and publish their work, but also consumes large amounts of energy. We propose an approach for comparing the performance of two algorithms: by performing runs on carefully chosen instances, we obtain a probabilistic statement on which algorithm performs best, trading off between the computational cost of running algorithms and the confidence in the result. We describe a set of methods for this purpose and evaluate their efficacy on diverse datasets from SAT, CSP and BNSL. On all these datasets, most of our approaches were able to choose the correct algorithm with about 95% accuracy, while using less than a third of the CPU time required for a full comparison; the best methods reach this level of accuracy within less than 15% of the CPU time for a full comparison. 2012 ACM Subject Classification General and reference → Evaluation; Theory of computation → Automated reasoning; Theory of computation → Constraint and logic programming