Evaluation of Estimation of Distribution Algorithm to Calibrate Computationally Intensive Hydrologic Model

Abstract

The estimation of distribution algorithm (EDA) is a new evolutionary algorithm developed as an alternative to the traditional genetic algorithm (GA). The EDA guides the search by avoiding the crossover and mutation operators of the GA in favor of building and sampling probabilistic distributions of promising candidate solutions. By increasing the probability of generating solutions with better fitness values, the EDA locates the region of the global optimum or its accurate approximation. In this study, EDA was used to calibrate the parameters of the soil and water assessment tool hydrologic model for the Xunhe River Basin in China. The EDA was compared with three other algorithms: (1) the Multistart Local Metric Stochastic Radial Basis Function algorithm (a surrogate optimization method), (2) the Shuffled Complex Evolution algorithm, and (3) the GA. Four metrics are presented to assess the performance of the algorithms: (1) efficiency in terms of the average best objective function value in a limited number of function evaluations, (2) variability in terms of standard deviation and the box plot, (3) reliability in terms of the empirical cumulative distribution function, and (4) accuracy in terms of the Nash–Sutcliffe efficiency coefficient and overall volume error. Results indicated that the EDA is more efficient and could provide more accurate solutions with a relatively high probability, at least for this case study.