Impact of Gaussian learning rate on training of sigmoidal FFANN using zero and random weight initializations

Artificial Neural Networks has number of applications in the various fields like filtering, pattern recognition, function approximation and control problems. Number of non-linear transformation problems can be solved using these fields. Randomizing the various parameters is a common paradigm like random initialization of weights to zero and generating weights randomly. To minimize the error the first order or second order derivatives are used. The training of the neural network is done so that the error is minimized. In this paper the artificial neural network parameter weights are initialized to zero and are generated randomly during the training of the network. Then the error obtained is compared when there is no change in the back-propagation algorithm and when changes are done with weights zero and random both. The results indicate that it is feasible to perform training by initializing the weights to zero also though the results are higher approximately by a factor of two or three from the results obtained when the weights are generated randomly and learning rates made Gaussian.