AutoML.org & IOHprofiler, featuring the switching squirrel. After fixing a minor bug, our Switching Squirrel achieved a strong 3rd place at the black-box challenge at #NeurIPS2020 (#AutoML and Hyperparameter Optimization). On the warm-start friendly leaderboard, our Squirrel was even ranked 1st and was awarded an extra prize money. Our Squirrel combines ideas from (i) (simple) meta-learning by recognizing known configuration spaces, (ii) a diverse portfolio of acquisition functions, surrogate models and data transformations for Bayesian Optimization and (iii) differential evolution for a greedy final stage. It is the outcome of a joint collaboration between the groups from AutoML.org (tnt@Leibniz University Hannover and University of Freiburg) and IOHProfiler (Leiden University and Sorbonne University).
Competition homepage