Model trust region technique in parallel Newton´s method for training feedforward neural networks

The double dogleg trust region approach of unconstrained minimization is introduced in the parallel Newton´s (PN) algorithm proposed by M. D. Zhao (1993). The PN algorithm uses a recursive procedure for computing both the Hessian matrix and the Newton direction. The input weights of each neuron in the network are updated after each presentation of the training data with a global strategy. Experimental results indicate that the double dogleg trust region approach is superior to the line search technique in the PN algorithm, and that the PN algorithm with both global strategies exhibits better convergence performance than the well-known backpropagation algorithm