Author(s): Youwei Qin; George Kuczera; Dmitri Kavetski
Linked Author(s):
Keywords: Robust Gauss-Newton algorithm; SCE; Model calibration; Snow model; SIMHYD model
Abstract: Parameter estimation remains an ongoing challenge in hydrological modeling for a number of reasons. One severe challenge is that of numerically rough and discontinuous objective function surfaces, which cause premature convergence of standard gradient-based methods. This difficulty has motivated development of gradient-free stochastic search methods such as Shuffled Complex Evolution (SCE). Stochastic optimization is generally more robust, but requires substantially more function evaluations. This paper introduces a robust Gauss-Newton method which can more reliably search over rough objective function surfaces. It uses a curve fitting technique to smooth the elements of the Jacobian matrix and an adaptive inexact line search method to update the finite difference perturbations used in the algorithm. A case study involving two conceptual hydrological models demonstrates the performance of the robust Gauss-Newton algorithm by comparing its convergence rate and computational cost with the SCE method. The results show the robust Gauss-Newton method can achieve comparable accuracy to the SCE algorithm at a much lower cost, with up to 4-20 times fewer objective function evaluations.
Year: 2013