This topic deals with the application of advanced GO methods to optimally choose hyperparameters in machine learning training. The idea is to use GO methods to find good hyperparameters based on a small set of trials, each consisting in chosing hyperparametes, training a network, evaluating its performance on a validation set. As a GO problem, this is black box, noisy, expensive. We (Matteo Lapucci, Fabio Schoen, Alessio Sortino) recently submitted a paper on this topic and we would like to experiment with machine learning.

Hyperparameter optimization in machine learning via black-box global optimization