A Novel Radial Basis Function (RBF) Network for Bayesian Optimization

Abstract

Gaussian process (GP) is the most popular surrogate model used in Bayesian optimization for solving computationally expensive problems. However, the computation time for constructing GP may become excessively long when the number of training samples increases. This study investigates multi-task learning with the radial basis function (RBF) network and proposes a multi-task learning network models based on RBF. Moreover, the proposed multi-task-RBF networks are applied to a Bayesian optimization framework and used to replace the GP for avoiding the covariance calculation. Experimental studies under several scenarios indicate that the proposed algorithm is competitive in performance compared with GP- and single-task-based Bayesian optimizations.