Author(s): Saman Razavi; Bryan A. Tolson

Linked Author(s):

Keywords: No Keywords

Abstract: Neural networks can act as highly flexible meta-models of response surface functions of water resources models. A wide variety of optimization algorithms have been used in the literature to train neural networks including local derivative-based methods, variations of back-propagation, and different stochastic global optimizations. Stochastic global optimization algorithms usually suffer from their high computational demand for neural network training. In this study, a recently developed computationally efficient global optimization algorithm called Dynamically Dimensioned Search (DDS) is utilized for neural network training. DDS is modified to perform a more intelligent search on the basis of the known internal behaviour of neural networks. This modified DDS is called neural network-based dynamically dimensioned search (NNDDS) algorithm. The performance of DDS and NNDDS are tested over a simple test function as well as on the meta-modeling of the response surface function of a MESH hydrologic model with 24 parameters. Moreover, a genetic algorithm as well as the Levenberg-Marquardt (LM) backpropagation method are also applied to the ANN training problems for comparison purposes. Results demonstrate that DDS performs better than GA. Although the new NNDDS algorithm easily outperforms both DDS and GA, LM is dominant in terms of training error minimization and convergence speed. However, preliminary validation results suggest that NNDDS may enhance the generalization capability of neural networks relative to LM.

DOI:

Year: 2009