B. Beliczynski et al. (Eds.): ICANNGA 2007, Part I, LNCS 4431, pp. 276–285, 2007.
© Springer-Verlag Berlin Heidelberg 2007
Automatic Design of ANNs by Means of GP for Data
Mining Tasks: Iris Flower Classification Problem
Daniel Rivero, Juan Rabuñal, Julián Dorado, and Alejandro Pazos
Department of Information & Communications Technologies, Campus Elviña
15071, A Coruña, Spain
{drivero,julian,juanra,apazos}@udc.es
http://sabia.tic.udc.es
Abstract. This paper describes a new technique for automatically developing
Artificial Neural Networks (ANNs) by means of an Evolutionary Computation
(EC) tool, called Genetic Programming (GP). This paper also describes a prac-
tical application in the field of Data Mining. This application is the Iris flower
classification problem. This problem has already been extensively studied with
other techniques, and therefore this allows the comparison with other tools. Re-
sults show how this technique improves the results obtained with other tech-
niques. Moreover, the obtained networks are simpler than the existing ones,
with a lower number of hidden neurons and connections, and the additional ad-
vantage that there has been a discrimination of the input variables. As it is ex-
plained in the text, this variable discrimination gives new knowledge to the
problem, since now it is possible to know which variables are important to
achieve good results.
1 Introduction
ANNs are learning systems that have solved a large amount of complex problems
related to different disciplines (classification, clustering, regression, etc.) [1]. The
interesting characteristics of this powerful technique have induced its use by research-
ers in different environments [2].
Nevertheless, the use of ANNs has some problems mainly related to their devel-
opment process. This process can be divided into two parts: architecture development
and training and validation. As the network architecture is problem-dependant, the
design process of this architecture used to be manually performed, meaning that the
expert had to test different architectures and train them until finding the one that
achieved best results after the training process. The manual nature of the described
process determines its slow performance although the recent use of ANNs creation
techniques have contributed to achieve a more automatic procedure.
The technique described in this paper allows the automatically obtaining of ANNs
with no need of human participation. This technique is applied to a well-known prob-
lem: the Iris Flower classification Problem [3]. Results show how this technique can
solve the problem. Moreover, it is capable of obtaining simpler networks than the
existing ones for solving this problem.