Geoscience Reference
In-Depth Information
IMPROVED ROBUSTNESS AND EFFICIENCY OF THE
SCE-UA MODEL-CALIBRATING ALGORITHM
NITIN MUTTIL
∗,§
and SHIE-YUI LIONG
†,¶
∗
Department of Civil and Structural Engineering
Hong Kong Polytechnic University, Hong Kong
§
cenitinm@polyu.edu.hk
†
Tropical Marine Science Institute
National University of Singapore, Singapore
¶
tmslsy@nus.edu.sg
The Shu
ed Complex Evolution (SCE-UA) has been used extensively and
proved to be a robust and e
cient global optimization method for the calibra-
tion of conceptual models. In this study, we propose two enhancements to the
SCE-UA algorithm to improve its exploration and exploitation of the search
space. A strategically located initial population is used to improve the explo-
ration capability and a modification to the downhill simplex search method
enhances its exploitation capability. This enhanced version of SCE-UA is tested
on a suite of test functions and it is observed that the strategically located ini-
tial population drastically reduces the number of failures and the modified
simplex search also leads to significant reduction in the number of function
evaluations to reach the global optimum, when compared to the original SCE-
UA. Thus, the two enhancements further improve the robustness and e
ciency
of the SCE-UA algorithm.
1. Introduction
With the advent of digital computers, a generation of models known as con-
ceptual models has been developed. The successful application of a model
heavily depends on how well it is calibrated. There is a substantial body
of research documenting problems encountered during model calibration,
especially with conceptual models.
1
-
3
Duan
1
pointed out five character-
istics that complicate the optimization of conceptual models. The most
important of these characteristics is the presence of multiple optima.
To deal with the problem of multiple local minima, global search meth-
ods are applied. These methods are global in the sense that they consti-
tute a parallel search of the search space (as opposed to a point by point
search) by using a population of potential solutions. This capability of such
techniques for effective “exploration” of the search space makes them less
291
