Effect of altering the Gaussian function receptive field width in RBF neural networks on aluminium fluoride prediction in industrial reduction cells

Artificial neural networks are increasingly useful computational models, consisting of highly interconnected parallel processing units. In particular, radial basis function, RBF, networks are emerging as important computational models for a broad range of applications. The Gaussian function used in RBF networks has an adjustable parameter, σ, which specifies the diameter of the receptive field of the hidden layer neurons. The selection of σ is commonly carried out using heuristic techniques. The selection of σ, as shown in this paper, plays an important role in the predictive capabilities of the RBF network. However, the use of a Gaussian function with the standard deviation of the training pattern output vector is shown to be associated with the minimum RMS error obtained using an optimum σ value derived using a heuristic technique. The aluminium fluoride, AlF ₃, content of industrial reduction cell for aluminium production is well predicted using the RBF network with a Gaussian function σ value derived using the standard deviation of the training pattern output vector