View article

In this work we propose to use meta-learning to learn sets of symbolic default hyperparameter configurations that work well across many data sets. A well known example for such a symbolic default is the logarithmic relation between the number of features of a dataset and the available features per split of a Random Forest, as observed by Breiman (2001). Symbolic functions allow for a more rich vocabulary to define defaults on. In the past, symbolic and static default values have been obtained either from hand-crafted heuristics or empirical evaluations of specific algorithms. We propose to automatically learn such symbolic configurations, ie, formulas containing meta-features, from a large set of prior evaluations of numeric hyperparameters on multiple data sets via symbolic regression and optimization.