AppART: An ART Hybrid Stable Learning Neural Network for Universal Function Approximation

Function Approximation with ARTMAP Architectures

“…Comparing the MSE BA score, obtained by single epoch training, and those reported in [14], where multi-epoch training was used, we can state that the BA clearly performs better.…”

“…Table 2 contains the results for MLP, RBF, GRNN, FAM, GAM, PROBART, FasBack, AppART and BA. For the first eight neural networks the results are from [14]. BA produces a very good MSE score for this regression task, most likely due to the optimized parameter values obtained by trial and error.…”

“…In the second test, we use the fifth-order chirp function [14]: g(x) = 0.5+0.5 sin(40πx 5 ). Marti et al have experimentally compared the function approximation performance of the following neural models [14]: AppART, Multi Layer Perceptron (MLP), RBF, General Regression Neural Network (GRNN), FAM, GAM, PROBART, and FasBack [23].…”

“…Marti et al have experimentally compared the function approximation performance of the following neural models [14]: AppART, Multi Layer Perceptron (MLP), RBF, General Regression Neural Network (GRNN), FAM, GAM, PROBART, and FasBack [23]. The reported score was the mean squared error (MSE).…”

“…AppART [14] is an ART-based neural network model that incrementally approximates continuousvalued multidimensional functions through a higher-order Nadaraya-Watson regression.…”