| Popis: |
Selection of input variables of the empirical models has vital effect on the prediction performance, reduced overfitting and reduced computational load. Various trials and error and sequential methods in the literature to deal with input selection for artificial neural networks (ANNs). However, these methods are not considered as automatic and systematic. This study proposes a novel and efficient mixed integer nonlinear programming-based approach to handle optimal input selection and the ANN training simultaneously for classification problems. Such selection uses binary (0-1) variables to represent the presence of the input variables and trains traditional continuous network weights simultaneously. Two classification case studies are given to demonstrate the advantages by using widely used data sets and statistical measures. The first data set is related to the characterization of the type of a tumor related to breast cancer, the second data set is about predicting the type of a biotechnological product using different features, the last one is related to heart failure prediction. Results show that better test performance can be achieved with optimally selected inputs, resulting in reduced overfitting. The proposed approach delivers a significant advantage during the design and training of the ANNs and is also applicable to other empirical models. |
