A0646
Title: Innovation in high-dimensional categorical Bayesian optimization
Authors: Timothee Bacri - University of Exeter (United Kingdom) [presenting]
Daniel Williamson - University of Exeter (United Kingdom)
Bertrand Nortier - University of Exeter (United Kingdom)
Abstract: Bayesian optimization is a powerful tool to optimize black-box and expensive objective functions. Handling high-dimensional inputs is challenging, but can be managed with methods such as embeddings. This allows the usage of traditional Gaussian processes as surrogates. A high-dimensional categorical optimization problem is looked into, and hence, discretization is required. However, the popular discretization method scaling-and-rounding suffers when dealing with large numbers of categories due to the non-uniform distribution of the input variables. An adaptation is proposed using quantile-binning instead. Bins being defined with quantiles means each category is equally likely over each variable, ensuring a balanced exploration over the high-dimensional categorical space. This method is evaluated, and its strengths and limitations are highlighted.
