Adaptive random search evaluated as a method for calibration of the SMA-NWSFS model

Abstract

Random search methods are becoming more widely used to estimate model parameters. Their ability to globally search a parameter space makes them attractive for solving problems that have multi-local optima, as are non-linear hydrologic models such as Conceptual Rainfall-Runoff (CRR) models. The investigation of this thesis is on the ability of the Adaptive Random Search (ARS) to find the global optimum of the CRR model known as the Soil Moisture Accounting Model of the National Weather Service River Forecast System (SMA-NWSRFS) and compares its performance to that of the Uniform Random Search (URS). Research results indicate that, although the ARS was slightly more efficient than the URS, neither strategy demonstrated an ability to converge to the globally optimum parameter set. Factors which inhibit the convergence include model structure characteristics and an insufficient number of points searched. Ways for random search techniques to identify and address these problems are discussed.