An Integrated ANN-GA Approach to Data Classification

In this paper, we present an advanced approach to data classification based on the integration of artificial neural networks (ANNs) and genetic algorithms (GAs). We modify neural network architecture in a two-stage process. During the first stage, GA finds a suboptimal neural network architecture: number of nodes, training algorithm, learning rate, etc. Then, the fitting of weight coefficients and bias is carried out in order to minimize GA fitness function. In final section of the paper, we compare the results of the conventional and the proposed approaches. Gradient methods generally used for ANN training can lead to the penetration of the functional of training quality in local minima. Lately GAs are often used to increase the quality of ANN training. GAs perform tasks of artificial evolution of main features of neural network structure in order to increase classification accuracy (2). In (3), two possible variants of ANN architecture optimization are described. In the first case, the chromosomes have the same length equal to the number of ANN layers, and each gene characterizes the number of nodes in the corresponding layer. If the layer is absent then in the corresponding locus the zero gene is prescribed. Another approach assumes that chromosomes have unequal length, and the first locus contains the gene responsible for the number of ANN layers, and following genes contain data about the number of nodes. In this case, the null values are not entered into the chromosome. So the chromosomes have unequal length. In order to perform the operation of crossing-over in the second case the chromosome with shorter length is modified zero genes occupying random loci in the range between the third and the penultimate loci.