An experimental comparison of the Bayesian Ying-Yang criteria and cross validation on experts number selection in original and alternative model for mixture of experts

Mixture of Experts (ME) and its alternative extension are very popular in supervised learning family. The Expectation-Maximization (EM) algorithm is a good iterative approach in maximum likelihood parameter estimation for the ME learning. Recently, a new criteria on selecting the correct number of experts for both the original and alternative model for ME are proposed by one of the present author, based on the so-called Bayesian Ying-Yang (BW) Learning Theory. In this paper, the new criteria is experimental studied and compared with the well known Cross Validation criteria. Simulation results show that the expert number obtained by the BW criteria is highly consistence with the minimum generalization error and outperforms Cross Validation approach.