HYBRIDIZATION OF ARTIFICIAL NEURAL NETWORK USING DESIRABILITY FUNCTIONS FOR PROCESS OPTIMIZATION

As desirability functions, proposed by many authors, follow most of the properties of standard transfer f unctions used for ANN, the objective of hybridsation in this study is to make use the property of desirability function in the neural network architecture and evaluate their performances while training and optimizing the architecture for an input- output relationship including the concept of composite desirability optimization technique when multiple responses are present. Two important desirability functions, proposed by Harrington, 1965 and Gatza et al., 1972 are used in different combinations with the most useful tan-hyperbolic transfer function using real life data. Three useful hybrid combinations of transfer/desirability functions are observed based on consistent simulation performance, number of nodes and a new measure of composite MSE is proposed here. The work on incorporating the knowledge of composite desirability into ANN architecture and exploiting the non-linearity in inputs versus outputs during normalization is also attempted.