COMBINED TRUST REGION METHODS FOR NONLINEAR LEAST-SQUARES

Trust region realizations of the Gauss-Newton method are commonly used for obtaining solution of nonlinear least squares problems. We propos e three efficient algorithms which improve standard trust region techn iques: multiple dog-leg strategy for dense problems and two combined c onjugate gradient Lanczos strategies for sparse problems. Efficiency o f these methods is demonstrated by extensive numerical experiments.