UTILIZING METAHEURISTICS TO GUIDE THE TRAINING OF NEURAL NETWORKS

Abstract

Neural networks (NN), have become increasingly popular due to their practical applications. NN training is a crucial stage in constructing a reliable model that can accurately predict data. The goal of NN training is to determine the best internal parameters to optimize the network's performance on test data, according to a specific metric. In this study, we explore the use of metaheuristics to guide the entire training process. Our approach involves identifying favorable areas of the search space and invoking an optimizer to intensify the search in these regions. To train NN, we implemented two metaheuristics, Variable Neighborhood Search (VNS) and the Memetic algorithm (MA), and measured their effectiveness using classification accuracy as an evaluation metric on publicly available classification datasets. The obtained results suggest that MA is able to outperform both VNS and traditional training methods.