Bayesian optimization package

Overview

We present an open source Python package, a Bayesian Optimizer model that uses an ARD Matérn 3/2 Kernel and Expected Improvement as the acquisition function.

Similarly to Spearmint, initial candidates are drawn from a Sobol sequence, then a subset of points with the highest acquisition score get optimized using L-BFGS-B.

With each new trial, the kernel parameters are optimized with respect to the Gaussian process: a multiplying constant, and the characteristic length scale of each dimension (defining diagonal covariances). The prior mean of the Gaussian process is assumed to be constant and zero.