CLASSIFICATION USING MULTI-MODEL CO-EVOLUTIONARY ENSEMBLES

摘要

Evolutionary neural networks and ensembles have been widely used in multiple domains for effective machine learning with strong generalizing ability. Large dimensionality is a major problem associated with these systems. Co-evolution involves cooperation amongst the participating individuals of the evolutionary process that results in better optimization. Research so far is focused upon the formulation of good evolutionary techniques to evolve neural networks, sometimes in modular or ensemble architecture. Here we propose a novel concept using numerous occurrences of two models of neural network in an ensemble architecture, whose outputs integrate using a probabilistic sum rule to give the final output of the system. The selection of the number of experts from these modules is made adaptive by an evolutionary approach. As a result the system not only optimizes the individual prospective experts, each of which denotes a multi-layer perceptron or a radial basis function neural network; we also carry forward an optimal selection of these experts. Experimental results on Breast Cancer disease prove that the algorithm can effectively learn and generalize. By comparing the proposed method with other methods in literature we find that the proposed algorithm has higher generalization and learning ability.