A new method in determining initial weights of feedforward neural networks for training enhancement

Abstract An algorithm for determining the optimal initial weights of feedforward neural networks based on a linear algebraic method is developed. The optimal initial weights are evaluated by using a least squares method at each layer. With the optimal initial weights determined, the initial error is substantially smaller and therefore the number of iterations required to achieve the error criterion is reduced. For a character recognition task, the number of iterations required for the network started with the optimal weights is only 53.9% of that started with the random weights. In addition, the time required for the initialisation process is negligible when compared to the training process.