Constructive Methods for Neural Network Ensembles Based on Two Selective Methods Matching

Twelve methods for constructing two-level selective neural network ensembles are investigated, as are realized by using two selective methods that accomplish two tasks of individual selection respectively. The effects of different corporations of selection methods and individuals' size on two-level selective neural network ensembles are discussed. Experimental results showed that the matching modes of methods of selection and proper individuals' number adopted can ensure the accuracy and diversity of individuals and alleviate the effect of the over-fitting problem. As a consequence, the accuracy of neural network ensembles is improved.