A computational methodology applied to optimize the performance of a river model under uncertainty conditions

Advances in computational science have made an explosion of computational models for analyzing and predicting the behavior of complex environmental systems possible, such as river models. Model accuracy is highly influenced by many sources of uncertainties, and one of these sources is parameter uncertainty. In this research, we present a search and optimization methodology to achieve a higher prediction quality of a computational system for calculating the translation of waves in rivers. Our proposal aims to achieve this goal using the least amount of computational resources. We address this issue by performing a two-phase optimization via simulation methodology. The first phase consists in a global exploration step over the entire search space. This phase identifies promising regions for optimization based on a neighborhood structure of the problem, using a Monte Carlo heuristic plus the K-means method. The second phase is a fine-grained approach that consists in seeking the best solution, either the optimum or a sub-optimum by performing a “reduced exhaustive search” in such promising regions. We achieve a speed-up of 20× when searching the best parameter settings in comparison with an exhaustive search in the whole space of candidates’ parameter settings. This acceleration is measured in terms of the number of simulations run required to find a solution. When using our methodology and parallel computing, we reduced from 11 to 0.5 days the complete time. We achieved a 22× gain, fulfilling the objective of reducing the use of computing resources.