Comparing Particle Swarm Optimization Approaches for Training Multi-Layer Perceptron Neural Networks for Forecasting

Abstract

Multilayer Perceptron Artificial Neural Networks (MLP-NN) have been widely used to tackle forecasting problems. The most used algorithm for training MLP-NN is called Backpropagation (BP). Since the BP presents a high chance to be trapped in local minima during the training process for forecasting, we propose in this paper to assess some recently proposed variations of the Particle Swarm Optimization algorithm (PSO) applied for this purpose. We tested the standard PSO, the APSO, the ClanPSO and the ClanAPSO in five benchmark data sets. Although the standard version of the PSO presented worse results when compared to the BP algorithm, we observed that the ClanAPSO outperformed the BP algorithm in most of cases.