Automated Algorithm Configuration and Parameter Tuning

The use of automated configuration and parameter tuning techniques plays an increasingly important role in the design, evaluation and application of high-performance algorithms for difficult computational problems. This chapter provides an introduction to these techniques and gives an overview of three families of procedures for automatically optimising the performance of parameterised algorithms. Racing procedures iteratively evaluate parameter settings on problem instances from a given set and use statistical hypothesis tests to eliminate candidate configurations that are significantly outperformed by other configurations. ParamILS uses a powerful stochastic local search method to search within potentially vast spaces of candidate configurations of a given algorithm. And finally, sequential model-based optimisation (SMBO) methods build a response surface model that relates parameter settings to performance, and use this model to iteratively identify promising settings. We also briefly survey other algorithm configuration and parameter tuning procedures, as well as related approaches, such as instance-based algorithm selection and configuration.