ADANN: Automatic Design of Artificial Neural Networks

Abstract

In this work an improvement of an initial approach to design Artificial Neural Networks to forecast Time Series is tackled, and the automatic process to design Artificial Neural Networks is carried out by a Genetic Algorithm. A key issue for these kinds of approaches is what information is included in the chromosome that represents an Artificial Neural Network. In this approach new information will be included into the chromosome so it will be possible to compare these results with those obtained in a previous approach. There are two principal ideas to take into account: first, the chromosome contains information about parameters of the topology, architecture, learning parameters, etc. of the Artificial Neural Network, i.e. Direct Encoding Scheme; second, the chromosome contains the necessary information so that a constructive method gives rise to an Artificial Neural Network topology (or architecture), i.e. Indirect Encoding Scheme. The results for a Direct Encoding Scheme (in order to compare with Indirect Encoding Schemes developed in future works) to design Artificial Neural Networks to forecast Time Series are shown.