An Entropy Approach to Evaluation Relaxation for Bayesian Optimization Algorithm

Bayesian Optimization Algorithm (BOA), a multivariate estimation of distribution algorithm, needs incorporating with efficiency enhancement techniques to be capable of solving difficult large-scale problems in a reliable and scalable manner. In this paper, we present a novel evaluation relaxation method which is based on the conditional entropy measurement. The concept of conditional entropy is rigorously analyzed and then is used to investigate the stability of the population. Especially, we utilize the evaluation relaxation strategy (ERS) proposed herein to determine whether a candidate solution should be evaluated by actual functions or be estimated by surrogate models. BOA coupled with our entropy-based ERS, termed en-BOA, shows its superiority in significantly reducing the total number of expensive fitness evaluations until reliable convergence. Experimental results prove that the entropy-based ERS enhances the efficiency of BOA while not negatively affecting the scalability of the original algorithm. In addition, the effects of our efficiency enhancement technique on population sizing requirements are also discussed.