Dynamic programming approach for multilayer neural network optimization

An algorithm for weight adjustments in a multilayer neural network is derived using the principles of dynamic programming. According to this algorithm at every layer the cost obtained using optimal values for weights for the remainder of the network is minimized using the weights in the current layer. Optimum weights are then computed recursively starting from the output layer. The mathematical formulation of the weight minimization problem is described within the dynamic programming framework. Equations governing weight adjustments in each layer are derived when the activation functions for neurons are continuous, e.g., sigmoid functions.<
>