Optimization of back-propagation network using simulated annealing approach

The back-propagation network (BPN) is a popular data mining technique. Nevertheless, different problems may require different network architectures and parameters. Therefore, rule of thumb or "try and error" methods are usually used to determine them. However, these methods may lead worse network architectures and parameters. A dataset may contain many features; however, not all features are beneficial for classification in BPN. Therefore, a simulated annealing (SA) approach is proposed to select the beneficial subset of features and to obtain the better network architectures and parameters which result in a better classification. In order to verify the developed approach, three dataset, namely PIMA, IONOS, and CANCER from UCI (University of California, Irvine) machine learning database, are employed for evaluation, and the 10-fold cross-validation is applied to calculate the classification result. Compared with the MONNA (Multiple ordinate neural network architecture) structure developed by Leazoray and Cardot, the classification accurate rates of the developed approach are superior to those of the MONNA. When the feature selection is taken into consideration, the classification accurate rates of three dataset are increased. Therefore, the developed approach can be utilized to find out the network architecture and parameters of BPN, and discover the useful attributes effectively.