Differential Evolution-based Neural Network Training Incorporating a Centroid-based Strategy and Dynamic Opposition-based Learning

Training multi-layer neural networks (MLNNs), a challenging task, involves\nfinding appropriate weights and biases. MLNN training is important since the\nperformance of MLNNs is mainly dependent on these network parameters. However,\nconventional algorithms such as gradient-based methods, while extensively used\nfor MLNN training, suffer from drawbacks such as a tendency to getting stuck in\nlocal optima. Population-based metaheuristic algorithms can be used to overcome\nthese problems. In this paper, we propose a novel MLNN training algorithm,\nCenDE-DOBL, that is based on differential evolution (DE), a centroid-based\nstrategy (Cen-S), and dynamic opposition-based learning (DOBL). The Cen-S\napproach employs the centroid of the best individuals as a member of\npopulation, while other members are updated using standard crossover and\nmutation operators. This improves exploitation since the new member is obtained\nbased on the best individuals, while the employed DOBL strategy, which uses the\nopposite of an individual, leads to enhanced exploration. Our extensive\nexperiments compare CenDE-DOBL to 26 conventional and population-based\nalgorithms and confirm it to provide excellent MLNN training performance.\n