Genetic Algorithm Optimization for Selecting the Best Architecture of a Multi-Layer Perceptron Neural Network: A Credit Scoring Case

Networks are powerful tools for classification and regression, but it is difficult and time costly to determine the best architecture for a given problem. In this paper, Genetic Algorithms (GA) are used to optimize the architecture of a Multi-Layer Perceptron Neural Network (MLP) in SAS®, in order to improve the predictive power of the credit risk scorecards. The objective function to maximize is the ROC curve and the input variables are the number of hidden layers and units, activation function, use or not of bias and whether it will be a direct connection between the initial and the final layer. Results show that this method outperforms logistic regression and the default neural network architecture of SAS Enterprise Miner™. The predictive power of this method is similar to the Global Optimum but in a reasonable time.