Forward additive neural network models
Byung-Hyuk Ahn · 1996
Artificial neural networks have been successfully applied in pattern recognition and function approximation. In managerial decision making, pattern recognition has been used for credit rating, market segmentation, and prediction of bank failures. Function approximation has been used in forecasting and mapping of complex relations between sets of variables. Neural networks for pattern recognition and function approximation must be trained. Training is a nonlinear minimization problem. However, since the problem is apt to contain multiple local minima, a methodology is needed to increase the likelihood of finding the global minimum. Another problem is how to determine the size of the network. The generalization capability of a neural network depends on its size. Many researchers tried to resolve both network size determination and global optimization. The study developed Forward Additive Neural Network (FANN) models to meet both goals. FANN models increase the network size by adding a node at a time. Using a statistical test, FANN models determine the appropriate size of the network. With some heuristic features, FANN models can provide a very good solution for neural network training. The results of experiments with several benchmark problems show that FANN models are effective.