Parameter by Parameter Algorithm with Goal Programming Method for Neural Network Classifiers

A new training algorithm for neural network classifier, PBPGP algorithm, is presented in this paper, The input errors of the output layer and hidden layer are taken into account. In each iteration step, the parameters, including the weights and thresholds can be optimized directly, one by one with other variables fixed, Generally, when calculating the desired hidden targets, a linear equation set is needed. But when the determinant of the coefficient matrix turns to be zero, the solution is not uniquely. That results in the stalling problem. Furthermore, a truncation error is caused by the sigmoid function during the solving process. Based on the idea of goal programming technique, this paper proposes a new method to calculate the desired hidden targets. An achievement function is set up for the hidden targets. The model gives a satisfied solution under any conditions. Effectiveness of the proposed method is demonstrated by a mushroom classification problem