HOW TO CHOOSE AN APPROPRIATE TRANSFER-FUNCTION IN DESIGNING A SIMPLEST ANN TO SOLVE SPECIFIC PROBLEMS

How to incorporate certain constraints known a priori into the transfe r function to reduce the number of hidden units of an MFNN in solving some specific problems is discussed. The N-bit parity and symmetrical problems are used as examples and the derivation of the transfer funct ion based on the a priori knowledge of the input patterns is given. It is shown that by use of the method proposed, only one hidden unit is needed in solving the parity and symmetry test problems.