Evolutionary Multiobjective Ensemble Learning Based on Bayesian Feature Selection

This paper proposes to incorporate evolutionary multiobjective algorithm and Bayesian Automatic Relevance Determination (ARD) to automatically design and train ensemble. The algorithm determines almost all the parameters of ensemble automatically. Our algorithm adopts different feature subsets, selected by Bayesian ARD, to maintain accuracy and promote diversity among individual NNs in an ensemble. The multiobjective evaluation of the fitness of the networks encourages the networks with lower error rate and fewer features. The proposed algorithm is applied to several real-world classification problems and in all cases the performance of the method is better than the performance of other ensemble construction algorithms.