OASC-2017: *Zilla Submission


Chris Cameron, Holger H. Hoos, Kevin Leyton-Brown, Frank Hutter ;
Proceedings of the Open Algorithm Selection Challenge, PMLR 79:15-18, 2017.


*Zilla is a model-based approach for algorithm selection and the most recent iteration of the well-known SATzilla project. The new *Zilla system has increased flexibility for the user and is configurable to run with many machine learning models and alternatives for various pre/post processing steps (e.g., presolver selection, feature completion prediction, and solver subset selection). The main additions to our *Zilla pipeline are automated procedures for feature group selection, hyper-parameter tuning, and solver subsampling prior to model building. We submit two versions for the competition that are equivalent except for the choice of per-instance machine learning model. For our first submission, we use a weighted pairwise random forest classifier. For our second submission, we test an experimental approach that offline, builds a weighted pairwise random forest classifier and online, finds the nearest instances based on the average path lengths across trees and optimizes a schedule over those instances.

Related Material