Training feedforward neural network via multiobjective optimization model using non-smooth L(1/2) regularization

摘要

The paper presents a new approach to optimize the Multilayer Perceptron Neural Network (MLPNN), to deal with the generalization problem. As known, most supervised learning algorithms aim to minimize the training error. However, the mentioned methods, based only on error minimizing, may generate a solution with an insufficient generalization performance. This present work proposes a multiobjective modelling problem involving two objectives: accuracy and complexity since the learning problem is multiobjective by nature. The learning task is carried on by minimizing both objectives simultaneously, according to Pareto domination concept, using NSGAII (Non-dominated Sorting Genetic Algorithm II) as a solver. This method leads us to a set of solutions called Pareto front, being the optimal solutions set, the adequate MLPNN need to be extracted. We show empirically that the proposed method is capable of reducing the neural networks topology and improved generalization performance, in addition to a good classification rate compared to different methods. (C) 2020 Published by Elsevier B.V.