J.F. Martínez-Trinidad et al. (Eds.): CIARP 2006, LNCS 4225, pp. 389 – 398, 2006. © Springer-Verlag Berlin Heidelberg 2006 Training of Multilayer Perceptron Neural Networks by Using Cellular Genetic Algorithms M. Orozco-Monteagudo, A. Taboada-Crispí, and A. Del Toro-Almenares Center for Studies on Electronics and Information Technologies, Universidad Central de Las Villas, Carretera a Camajuaní, km 5 ½ , Santa Clara, VC, CP 58430, Cuba morozco@uclv.edu.cu, ataboada@uclv.edu.cu, anestodt@uclv.edu.cu Abstract. This paper deals with a method for training neural networks by using cellular genetic algorithms (CGA). This method was implemented as software, CGANN-Trainer, which was used to generate binary classifiers for recognition of patterns associated with breast cancer images in a multi-objective optimiza- tion problem. The results reached by the CGA with the Wisconsin Breast Can- cer Database, and the Wisconsin Diagnostic Breast Cancer Database, were compared with some other methods previously reported using the same data- bases, proving to be an interesting alternative. Keywords: Neural networks, genetic algorithms, cellular automata, multi- objective classification. 1 Introduction The trend of using Multilayer Perceptron Neural Networks (MLP) [1] for the solution of classification problems in pattern recognition applications is understandable due to their capacity to imitate the nature of the human brain (learning capacity), and the fact that their structure can be formulated mathematically. The functionality of the topol- ogy of the MLP is determined by a learning algorithm able to modify the parameters of the net. The algorithm of Backpropagation (BP), based on the method of steepest descent [1] in the process of upgrading the connection weights, is the most commonly used by the scientific community. The main limitations and problems that present the BP algorithm in training the MLP are exposed in [2]. Recently, numerous works have been reported trying to overcome their main limitations. However, the topology selec- tion issue [3] for MLP still leaves margin for improvements. At the same time, it is not clear what algorithm, or combination of algorithms, is the most appropriate to achieve the objective of reducing complexity of the classifiers and, simultaneously, increasing their benefits in terms of classification effectiveness, for a particular appli- cation, and much less for one with a more general character. On the other hand, Evolutionary Algorithms (EA) have demonstrated great effec- tiveness to solve problems of Multiobjective Optimization (MO) [4]. In addition, the use of Hierarchical Codification [5] (Fig. 1), combined with the EA, provides advan- tages in the determination of solutions for problems where the determination of a good structure (ignored a priori), is of vital importance. The enormous capacity of computation of the Cellular Automata (CA) [6] gives, in conjunction with EA, a very