Learning discrete mappings-Athena´s approach

The general problem of learning discrete mappings is considered. The focus is on the problem of learning a mapping that generalizes an incomplete description specified by a set of examples. The authors´ view of this problem´s nature is explained for the case of incompletely specified mappings, and on this basis a quantitative measure is given for determining the target learning performance. A neural-net model is proposed and shown to be appropriate for this general task. The model´s structure is briefly described and the principles and insight of an earlier proposed adaptive process are explained. How the target of this adaptive process relates to the goal of the adaptation, as the latter is specified, is explained. The model is qualitatively analyzed and compared with the multilayer perceptron and the nonlinear feedforward model. It is shown that the proposed model is functionally equivalent to the perceptron. A number of test examples are presented which provide an insight and an evaluation of the model´s performance