Applying dynamic surrogate models in noisy genetic algorithms to optimize groundwater remediation designs

Computational cost is a critical issue for large-scale water-resource optimization under uncertainty, since time-intensive Monte Carlo simulations are often required to evaluate over multiple parameter realizations. This paper presents an efficient approach for replacing most Monte Carlo simulations with surrogate models within a noisy genetic algorithm (GA). The surrogates are trained to predict the posterior expectations online on the basis of stochastic decision theory, using Monte Carlo simulation results created during the GA run. The surrogates, which in this application are neural networks, are adaptively updated to improve their prediction performance as the search progresses. A Latin hypercube sampling method is used to efficiently sample parameters for the Monte Carlo simulation, and the sampling results are archived so that the estimate of posterior expectation can be iteratively improved in an efficient manner. In addition, the GA is modified to incorporate hypothesis tests in its selection operator to account for sampling noise. The method is applied to a field-scale groundwater remediation design case study, whereas the primary source of uncertainty stems from hydraulic conductivity values in the aquifers. The results show that the method identified more reliable and cost-effective solutions with 86-90% less computational effort than the purely physically based noisy GA approach.