Number of hidden nodes for shape preserving ANN representation of a curve

Scientific knowledge is often available in handbooks and journals in the form of curves. ANN representation of curves is necessary for including them in the knowledge base of a connectionist expert system. Apart from one input node and one output node, representing abscissa and ordinate of the curve, respectively, the ANN has a hidden layer with s neurons, with sigmoid activation function. A method is developed to choose the value of s, logically, and for fast determination of weights and bias values, ensuring shape preservation of the curve. The method is applied to three examples for which the results of conventional techniques like backpropagation, RBF etc. are available in literature. The comparison brings out the advantages of the method developed