Domain-Dependent Parameter Selection of Search-Based Algorithms Compatible with User Performance Criteria

Abstract

Search-based algorithms, like planners, schedulers and satis-fiability solvers, are notorious for having numerous parame-ters with a wide choice of values that can affect their perfor-mance drastically. As a result, the users of these algorithms, who may not be search experts, spend a significant time in tuning the values of the parameters to get acceptable perfor-mance on their particular problem domains. In this paper, we present a learning-based approach for automatic tuning of search-based algorithms to help such users. The benefit of our methodology is that it handles diverse parameter types, per-forms effectively for a broad range of systematic as well as non-systematic search based solvers (the selected parameters could make the algorithms solve up to 100 % problems while the bad parameters would lead to none being solved), incor-porates user-specified performance criteria (φ) and is easy to implement. Moreover, the selected parameter will satisfy φ in the first try or the ranked candidates can be used along with φ to minimize the number of times the parameter settings need to be adjusted until a problem is solved.