Artificial neural networks (ANNs) have become common data driven tools for modeling complex, nonlinear problems in science and engineering. Many previous applications have relied on gradient-based search techniques, such as the back propagation (BP) algorithm, for ANN training. Such techniques, however, are highly susceptible to premature convergence to local optima and require a trial-and-error process for effective design of ANN architecture and connection weights. This paper investigates the use of evolutionary programming (EP), a robust search technique, and a hybrid EP–BP training algorithm for improved ANN design. Application results indicate that the EP–BP algorithm may limit the drawbacks of using local search algorithms alone and that the hybrid performs better than EP from the perspective of both training accuracy and efficiency. In addition, the resulting ANN is used to replace the hydrologic simulation component of a previously developed multiobjective decision support model for watershed management. Due to the efficiency of the trained ANN with respect to the traditional simulation model, the replacement reduced the overall computational time required to generate preferred watershed management policies by 75%. The reduction is likely to improve the practical utility of the management model from a typical user perspective. Moreover, the results reveal the potential role of properly trained ANNs in addressing computational demands of various problems without sacrificing the accuracy of solutions.


Civil and Environmental Engineering



URL: https://digitalcommons.calpoly.edu/cenv_fac/263