Block truncated-Newton methods for parallel optimization (Q582212)

The authors present a truncated Newton method for the minimization of a nonlinear function suitable for a parallel computer. In such a method a search direction is computed by approximately solving Newton equations using an iterative scheme. In order to make this method more suitable for parallel computation, the authors propose the use of a block iterative method based on the conjugate-gradient method. The authors claim that the advantage of such an approach is that a degree of parallelism is introduced in both the linear algebra and derivative calculations. Computational experience on an Intel hypercube computer is also reported.