Evolutionary ELMs with Alternative Treatments for the Population Out-Bounded Individuals

Abstract

Extreme Learning Machine (ELM) has been introduced as an algorithm for the training of Single-Hidden Layer Feedforward Neural Networks, capable of obtaining faster performances than traditional gradient-descendent approaches, such as Back-Propagation algorithm. Although effective, ELM suffers from some drawbacks, since the adopted strategy of random determination of the input weights and hidden biases may lead to non-optimal performances. Many Evolutionary Algorithms (EAs) have been employed to select input weights and hidden biases for ELM, generating Evolutionary Extreme Learning Machine (EELM) models. In this work, we evaluate the influence of three different treatments to handle the population out-bounded individuals problem in EAs by comparing three different Evolutionary Extreme Learning Machine approaches. The experimental evaluation is based on a rank system obtained by using Friedman hypothesis tests in relation to the experiments performed on ten benchmark data sets. The experimental results pointed out that some treatments to handle the out-bounded individuals are more adequate than others for the selected problems, and also, some EELMs are more sensible to the way that out-bounded individuals are treated than others.