Accuracy of neural network approximators in simulation-optimization

Heuristic search techniques are highly flexible but computationally intensive optimization methods that require hundreds, sometimes thousands, of evaluations of the objective function to reach termination criteria in common water resources optimization applications. One way to make these techniques more tractable when the objective function depends on a time-consuming flow and transport model is to employ an empirical approximation of the model. The current study examines the impact of employing artificial neural networks (ANNs) and linear approximators (LAs) on the quality and quantity of solutions obtained from simulated annealing-driven searches on two different ground-water remediation problems. The quality of results obtained when ANNs served as substitutes for the full model was consistently comparable to that of results obtained when the full model itself was called in the course of the search. The effect on quality of results of substituting an LA for the full model was more variable.