Automatic parameterization of neural networks using evolutionary algorithms

Abstract

Artificial Intelligence consists of an area where methods or systems are developed that act intelligently, approaching human behavior, in situations involving problem solving, acquisition and representation of knowledge, pattern recognition, etc. Within this context, a type of computational model that has gained prominence are the Artificial Neural Networks (ANNs), which are formed by basic blocks inspired by the biological neuron. An ANN has the ability to act in various applications, such as universal approximation of functions, process control, pattern recognition and classification, and data grouping. To meet a wide range of applications, an ANN requires the determination of a series of parameters, among them: topology, number of layers, number of neurons, activation function, training method, etc. In other words, the design of an ANN with the most appropriate configuration for each type of problem requires a series of choices, preliminary tests and experience from the designer. However, in order to avoid such choices being made empirically, it is possible to treat this parameterization as an optimization problem, allowing its resolution through the use of evolutionary algorithms, which are optimization tools developed to simulate several natural evolutionary processes. In this work, the Genetic Algorithm and Differential Evolution with binary coding were applied to automatically parameterize single-layer hidden neural networks applied in the modeling of a buck converter and in the prediction of compressive strength of self-compacting concrete (SCC) with the addition of fibers. The neural networks used were trained with the Extreme Learning Machine algorithm and the results of the simulations show that the Genetic Algorithm was the technique that presented the best performance when parameterizing the network in the modeling process of the buck converter, while the Differential Evolution combined with the binary coding GVP was the best strategy to parameterize the neural network in the process of predicting compressive strength of SCC.

DOI:https://doi.org/10.56238/sevened2024.007-089