On the Optimality Gap of Warm-Started Hyperparameter Optimization

Abstract

We study the general framework of warm-started hyperparameter optimization (HPO) where we have some source datasets (tasks) where we have already performed HPO, and we wish to leverage the results of these HPO to warm-start the HPO on an unseen target dataset (and perform few-shot HPO). Various meta-learning schemes have been proposed over the last decade (and more) for this problem. In this paper, we theoretically analyse the optimality gap of the hyperparameter obtained via such warm-started few-shot HPO, and provide novel results for multiple existing meta-learning schemes. We show how these results allow us identify situations where certain schemes have advantage over others.