The Use of Adaptive Logic Nets to Quantify Tar Sands Feed {draft{

The usefulness of Adaptive Logic Networks (ALNs) to analyse Near-Infrared (NIR) spectra of tar sands feedstock in a plant situation is examined. Laboratory experiments with tar sands do show good correlation between NIR and physical analysis using traditional mathematical and back propagation neural network techniques. The plant results are inconclusive due to poor correlation between the spectra and the physical grab samples. Regardless which of the three correlation techniques used the error of prediction is similar. A priori knowledge reduced the data set needed for input from 585 variables to 25. The reduction did not compromise the resolution of the spectra and is based on the shape of peaks. Training and testing of ALN trees are fast despite the large number of nodes. ALN trees can be easily converted to hardware which is needed in a harsh plant environment. More work is currently being done to reduce the error between the observed spectral results and the physical analysis. CONTENTS 1