Criteria for Best Architecture Selection in Artificial Neural Networks

Abstract

Architecture selection in artificial neural networks is a critical process which determines a satisfactory neural network model(s) that will lead to the most accurate results. The architecture that minimizes the difference between the target values of the neural network and the predictions produced by the model represents the best forecasts, namely the most appropriate model. In the literature, there are many common criteria for measuring model performance. In addition, some modified criteria, called weighted criteria, are suggested by combining the common criteria. In this study, the performances of the criteria available in the literature are compared by using both simulated and real-world datasets. We used three different exchange rate time series, four simulated time series with different structures and three well-known real-world datasets. The results show that the performances of the unweighted criteria vary depending on the data structure. However, the weighted criteria have performances as good as the popular criteria or better.