Improving estimation of distribution algorithms with heavy-tailed student's t distributions
Bin Liu, Shi Cheng, Yuhui Shi · 2018
As a derivative-free optimization method, the estimation of distribution algorithm (EDA) usually leverages a Gaussian or a Gaussian mixture model to represent the distribution of promising solutions that have been found so far. This paper investigates the application of an alternative model, namely the heavier-tailed Student's t distribution, to implement EDA. Two corresponding algorithms, termed ESTDA and EMSTDA, are developed, respectively. The ESTDA employs a single Student's t model to represent the distribution of the promising solutions. The EMSTDA uses a mixture of Student's t models to take account of hard multimodal cases. These methods are evaluated through extensive and in-depth numerical experiments using over a dozen of benchmark objective functions. Empirical results show that they provide remarkably better performance than their Gaussian counterparts in most cases under consideration.