Sundaram, Karthik.

This thesis addresses studies on cost-functions developed on the basis of maximum entropy principle, for applications in artificial neural network (ANN) optimization endeavors. The maximization of entropy refers to maximizing Shannon information pertinent to the difference in the output and the teacher value of an ANN. Apart from the Shannon format of the negative entropy formulation a set of Csiszar family functions are also considered. The error-measures obtained, via these maximum entropy formulations are adopted as cost-functions in the training and prediction schedules of a test perceptron. A comparative study is done on the performance of these cost-functions in facilitating the test network towards optimization so as to predict a standard teacher function sin (.). The study is also extended to predict a parameter (such as cell delay variation) in a practical ATM telecommunication system. Concluding remarks and scope for an extended study are also indicated.