Bayesian Optimisation for Objective Functions with Varying Smoothness

Abstract

Bayesian optimisation is a popular method in optimising complex, unknown and expensive objective functions. In complex design optimisation problems, the additional information about the smoothness, monotonicity or the modality of the unknown objective functions can be obtained either from the domain expertise or from the problem environment. Incorporating such additional information can potentially enhance the performance of the optimisation. We propose a methodology to incorporate the aforesaid extra information to have a better fitted surrogate model of the unknown objective function. Specifically, for Gaussian Process regression, we propose a covariance function to encompass varying smoothness across the input space through a parametric function whose parameters are tuned from the observations. Our experiments on both synthetic benchmark functions and real-world applications demonstrate that embodying such additional knowledge accelerates the convergence.