Título: Multi -objective Training of Multilayer Perceptron Using the Generalized Differential Evolution 3 Algorithm
Autores: Wesley Marques Lima, Igor Caetano Diniz & Honovan Paz Rocha
Resumo: This paper investigates the training of Multilayer Perceptron (MLP) neural networks using evolutionary algorithms. Initially, a classical Differential Evolution (DE) algorithm is used to train the MLP with a single-objective approach that minimizes the Mean Squared Error (MSE). The training is then reformulated as a multi-objective optimization problem, using the Generalized Differential Evolution 3 (GDE3) algorithm to simultaneously minimize the MSE and the norm of the weight vector. Experiments on five benchmark binary classification datasets demonstrate that the multi-objective approach generates smoother decision boundaries and more stable models compared to single-objective training. While both approaches achieve similar average accuracy, the GDE3-based method consistently exhibits lower variance and better generalization. We argue that Pareto-based multi-objective optimization offers practical advantages for neural network training, particularly in applications where model robustness and stability are critical, even at the cost of increased running time.
Palavras-chave: Multi-objective optimization; Differential Evolution; GDE3; neural networks; MLP; machine learning.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191807
Artigo em PDF: CBIC_2025_paper1191807.pdf
Arquivo BibTeX:
CBIC_2025_1191807.bib