Improving Generalization of Neural Networks Using MLP Discriminant Based on Multiple Classifiers Failures

Multiple classifier systems or ensemble is an idea that is relevant both to neural computing and to machine learning community. Different MCSs can be designed for creating classifier ensembles with different combination functions. However, the best MCS can only be determined by performance evaluation. In this study, MCS is used to construct discriminant set that was used to discriminate the difficult to learn from the easy to learn patterns. Hence, this study explores several potentially productive ways in which an appropriate discriminant set or failure treatment might be developed based on the selection of the two failure cases: training failures and test failures. The experiments presented in this paper illustrate the application of discrimination techniques using multilayer perceptron (MLP) discriminants to neural networks trained to solve supervised learning task such as the Launch Interceptor Condition 1 problem. The experimental results reveal that directed splitting using an MLP discriminant is an important strategy in improving generalization of the networks.