Designed Experiments for Tuning Hyperparameters in Machine Learning Algorithms

Most machine learning algorithms involve tuning a set of hyperparameters to achieve better performance. The most widely used methodologies for hyperparameter tuning are exhaustive grid search, random search and Bayesian optimisation. However, these methods can be time-consuming and may not achieve optimal outcomes as interactions between hyperparameters are ignored. We propose an alternative method for tuning machine learning hyperparameters through supersaturated designs and response surface methodology (SDRSM). In this talk, we demonstrate our approach SDRSM, and discuss the potential and limitation of applying the proposed method to hyperparameter search.

Nov 24, 2020 1:00 PM — Nov 25, 2020 3:00 PM