Hydro-informatics and Innovative Technology

Oral Abstract

Parallel Pareto Ranking Dynamically Dimensioned Search (PR-DDS) Algorithm for Efficient Parameter Estimation of Computationally Expensive Watershed Models

Wednesday, January 4
10:30 AM - 12:30 PM
Location: 5th Meeting Room

Computational intensity of distributed watershed models makes automatic calibration a computationally expensive process. The Dynamically Dimensioned Search (DDS) algorithm proposed by Tolson and Shoemaker (2007) is very effective in efficient parameter estimation / calibration of computationally expensive watershed models. This study introduces a parallel version of DDS, called PR-DDS. PR-DDS uses non-dominated sorting (also called Pareto ranking) of evaluated points to choose multiple points for simultaneous dynamically dimensioned search. The two objectives for Pareto ranking of evaluated points are 1) objective function value of a point, and 2) minimum distance of the point form other evaluated points. We test PR-DDS on the Cannonsville watershed model, developed in the Soil and Water Assessment Tool (SWAT). Performance of PR-DDS is tested with 1, 8, 32 and 64 processors and wall-clock time and computational speed-up are reported. Results indicate that PR-DDS performs considerably better (when parallel efficiency and speed-up are considered) than the serial DDS and is an effective parallel algorithm for efficient parameter estimation of computationally expensive watershed models.

Send Email for Taimoor Akhtar


Assets

Parallel Pareto Ranking Dynamically Dimensioned Search (PR-DDS) Algorithm for Efficient Parameter Estimation of Computationally Expensive Watershed Models



Attendees who have favorited this

Send Email for Parallel Pareto Ranking Dynamically Dimensioned Search (PR-DDS) Algorithm for Efficient Parameter Estimation of Computationally Expensive Watershed Models