Centroid-Based Differential Evolution with Composite Trial Vector Generation Strategies for Neural Network Training

Abstract

The learning process of feedforward neural networks, which determines suitable connection weights and biases, is a challenging machine learning problems and significantly impact how well neural networks work. Back-propagation, a gradient descent-based method, is one of the most popular learning algorithms, but tends to get stuck in local optima. Differential evolution (DE), a popular population-based metaheuristic algorithm, is an interesting alternative for tackling challenging optimisation problems. In this paper, we present Cen-CoDE, a centroid-based differential evolution algorithm with composite trial vector generation strategies and control parameters to train neural networks. Our algorithm encodes weights and biases into a candidate solution, employs a centroid-based strategy in three different ways to generate different trial vectors, while the objective function is based on classification error. In our experiments, we show Cen-CoDE to outperform other contemporary techniques.