Structure design of artificial neural networks using genetic algorithm

Artificial Neural Networks (ANN) have been successfully applied to various kinds of business and engineering problems, especially those involved in pattern classification. However, because of the lack of design standard or guidelines, the structure of specific ANN depends on the designer's own experiments or choices. In other words, even though we could construct a better ANN, we often settled down with just a satisfactory ANN. The purpose of this research is to apply the Genetic Algorithm (GA) to design a structure of ANN that yields better performance compared to the existing test results. For a bankruptcy prediction problem, an existing research using ANN which consists of 22 input processing elements (PEs) for financial ratios and 5 hidden PEs showed 70% hit ratio. In our research, the input financial ratios and the number of hidden PEs are determined by GA. The best ANN, which consists of 8 input PEs and 6 hidden PEs, shows 78.03% hit ratio. In addition, we compare the performance of two types of reproduction schemes, i.e., generational reproduction and steady-state reproduction.