Main content area

Evaluation of Dynamically Dimensioned Search Algorithm for Optimizing SWAT by Altering Sampling Distributions and Searching Range

Yen, Haw, Jeong, Jaehak, Smith, Douglas R.
JAWRA Journal of the American Water Resources Association 2016 v.52 no.2 pp. 443-455
Soil and Water Assessment Tool model, agricultural watersheds, algorithms, calibration, guidelines, normal distribution, prediction, sampling, simulation models, variance, water quality, Texas
The primary advantage of Dynamically Dimensioned Search algorithm (DDS) is that it outperforms many other optimization techniques in both convergence speed and the ability in searching for parameter sets that satisfy statistical guidelines while requiring only one algorithm parameter (perturbation factor) in the optimization process. Conventionally, a default value of 0.2 is used as perturbation factor in DDS where normal distribution is applied as sampling distribution with zero mean and the variance of one. However, the sensitivity of the perturbation factor to the performance of DDS in applications of watershed modeling is still unknown. The fixed-form sampling distribution may result in finding optimal parameters at the local scale rather than global scale in the parameter sampling space. In this study, the efficiency of DDS was evaluated by altering the perturbation factor (from 0.05 to 1.00) and the selection of sampling distribution (normal and uniform distribution) on hydrologic and water quality predictions in a lowland agricultural watershed in Texas, USA. Results show that the use of altered perturbation factors may cause variations in convergence speed or the ability in finding better solutions. In addition, DDS results were found to be very sensitive to sampling distribution selections. It is concluded that DDS-Normal outperformed DDS-Uniform in all scenarios. The choice of sampling distributions could be the potential major factor to be consiered for the performance of auto-calibration techniques in calibration problems for watershed simulation models.