Abstract
Many versions of Adaptive Resonance Theory (ART) have been developed. One of these is Distance based ART (DART), which can deal with both continuous and integer inputs, employs a dissimilarity based vigilance measure, and accept dynamically correlated data. However, DART performs only the clustering step. To include a model building step, we proposed two neural networks based on DART --- DART+RBFN and DART+GRNN. Using DART, a representative training set and a test set can be mined from a large set of data. These data can be used to build a radial basis function network (RBFN) or a generalized regression network (GRNN).DART+RBFN and DART+GRNN are tested using two SISO functions --- Gaussian curve function and trigonometric curve function, and two MISO functions --- Himmeblau function and Peaks function. We found that our method works well but DART+RBFN is better than DART+GRNN due to its superior modeling ability. Therefore DART+RBFN model was applied to construct an empirical model for Melt Index (MI) of a PE plant. We found that the model can be used to correlate and predict MI different steady operations as well as the changes of MI when there are grade transitions.