Feedforward neural networkʹs sensitivity to input data representation Original Research Article

Neural networks can be used to develop solutions to problems which are strictly symbolic. A question arises how to represent symbols in terms of number vectors understandable to neural networks. Data representation used should promote good generalization and reduce simulation uncertainty of the resulting model. Straightforward methods, which are most widely used, result in large networks which can prohibit solution of large problems. In the paper some new methods, which try to build information about the problem at hand into the representation, are proposed. It is shown that they are less sensitive to input data errors.