Modelos híbridos para construção de redes neurais artificiais via programação genética cartesiana

Abstract

In machine learning, the problem of data classification consists of correctly labeling unknown instances based on the labels of a set of known instances. An important method for solving data classification problems is called artificial neural networks (ANNs). ANNs are bioinspired computational methods, whose parameters must be correctly adjusted in order to solve a given learning task. Topology and weights are determining parameters for the performance of these methods. Despite the significance of ANNs in practical situations, and the several works available in the literature, to adjust its parameters remains as a current problem. Hence, the development of strategies to assist users during ANN modeling is relevant. One of these strategies is to use evolutionary algorithms (EAs) to optimize ANN parameters. The combination of ANNs and EAs is called neuroevolution. This work proposes new hybrid neuroevolutionary methods based on cartesian genetic programming (CGP) for the construction of ANNs. The methods carry out a decoupled optimization of the topology and the weights to generate the ANNs. Topology optimization is performed by CGP. For the weights optimization, two techniques are used separately: (i) differential evolution (DE), and (ii) backpropagation (BP). In the computational experiments these models were applied to eighteen data classification problems using benchmark datasets from the literature. There are experiments with balanced datasets and others with unbalanced datasets. The models were submitted to three performance studies. The first study evaluated the performance of the models using accuracy as the objective function. The second one evaluated the performance of the models using the mean square error as the objective function. The third study used the models to evaluate the performance of four distinct objective functions: accuracy, G-mean, Fβ-score, and area under the ROC curve. The results showed the superiority of the proposals when compared to alternative techniques from the literature.