AutoML with Monte Carlo Tree Search
Résumé
The sensitivity of machine learning (ML) algorithms w.r.t. their hyper-parameters and the difficulty of finding the ML algorithm and hyper-parameter setting best suited to a given dataset has led to the rapidly developing field of automated machine learning (AutoML), at the crossroad of meta-learning and structured optimization. Several international AutoML challenges have been organized since 2015, motivating the development of the Bayesian optimization-based approach Auto-Sklearn (Feurer et al., 2015) and the Bandit-based approach Hyperband (Li et al., 2016). In this paper, a new approach, called Monte Carlo Tree Search for Algorithm Configuration (Mosaic), is presented, fully exploiting the tree structure of the algorithm portfolio and hyper-parameter search space. Experiments (on 133 datasets of the OpenML repository) show that Mosaic performances match that of Auto-Sklearn.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...