Optimizing back-propagation networks via a calibrated heuristic algorithm with an orthogonal array

Improving the performance of neural networks is of considerable importance. Although previous studies have investigated how to design the optimal neural network, the heuristic algorithms developed to support the optimization process contain flaws. These heuristic algorithms do not perform efficiently and they require prior expert knowledge. This study commences by employing an orthogonal array using the Taguchi method to calibrate the factor levels of a heuristic algorithm and to estimate the percent contribution from various individual factors. Subsequently, the calibrated heuristic algorithm is used to optimize a back-propagation network (BPN). Changing the level of each individual factor systematically and then analyzing the main and interactive effects of the design factors by using the analysis of variance (ANOVA) leads to the optimal heuristic algorithm factor levels with regard to experimental cost. The proposed optimization procedure is demonstrated on the classification problems using the University of California’s Department of Information and Computer Science (ICS) server. The results indicate that the quality of the solution from the proposed approach is superior to that from a non-calibrated conventional design.