Student's t Distribution based Estimation of Distribution Algorithms for Derivative-free Global Optimization.

Bin Liu,Shi Cheng,Yuhui Shi
2016-01-01
Abstract:In this paper, we are concerned with a branch of evolutionary algorithms termed estimation of distribution (EDA), which has been successfully used to tackle derivative-free global optimization problems. For existent EDA algorithms, it is a common practice to use a Gaussian distribution or a mixture of Gaussian components to represent the statistical property of available promising solutions found so far. Observing that the Studentu0027s t distribution has heavier and longer tails than the Gaussian, which may be beneficial for exploring the solution space, we propose a novel EDA algorithm termed ESTDA, in which the Studentu0027s t distribution, rather than Gaussian, is employed. To address hard multimodal and deceptive problems, we extend ESTDA further by substituting a single Studentu0027s t distribution with a mixture of Studentu0027s t distributions. The resulting algorithm is named as estimation of mixture of Studentu0027s t distribution algorithm (EMSTDA). Both ESTDA and EMSTDA are evaluated through extensive and in-depth numerical experiments using over a dozen of benchmark objective functions. Empirical results demonstrate that the proposed algorithms provide remarkably better performance than their Gaussian counterparts.
What problem does this paper attempt to address?