Model-based optimization of algorithm parameters
By Frank Hutter
The task of setting an algorithm's parameters for optimal performance on a problem or a set of problems is ubiquitous in computer science. To motivate our problem domain, I will briefly review our previous success in applying model-free optimization techniques to configure algorithms with many, partly categorical parameters, such as CPLEX for mixed integer programming and both tree search and local search algorithms for SAT. Then, I will discuss our more recent work on model-based parameter optimization, which extends EGO, a classic algorithm for the global optimization of expensive blackbox functions. In particular, in order to deal with categorical parameters, I will introduce two novel approaches: a new kernel for Gaussian processes and an alternative predictive model based on random forests. I will also demonstrate a number of issues that arise in the presence of noisy function evaluations and when using transformations of the response variable to improve the model fit.
This talk is based on joint work with Holger Hoos, Kevin Leyton-Brown, and Kevin Murphy.