Comment on “How effective and efficient are multiobjective evolutionary algorithms at hydrologic model calibration?” by

In a recent paper by Tang, Reed and Wagener (2006, hereafter referred to as TRW) a comparison assessment was presented of three state-of-the-art evolutionary algorithms for multiobjective calibration of hydrologic models. Through three illustrative case studies, TRW demonstrate that the Strength Pareto Evolutionary Algorithm 2 (SPEA2) and Epsilon Dominance Nondominated Sorted Genetic Algorithm (ε-NSGAII) achieve a better performance than the Multiobjective Shuffled Complex Evolution Metropolis algorithm (MOSCEM-UA), previously developed by us and presented in Vrugt et al. (2003). I would like to congratulate TRW with their paper, which I believe provides a strong and valuable contribution to the field of hydrologic model calibration. However, I wish to differ in opinion about some of the main conclusions presented in their paper, especially with respect to the seemingly inferior performance of the MOSCEM-UA algorithm. The results presented in TRW were obtained using uniform random sampling of the initial parameter space. Such a sampling strategy is widely used within the water resources and computational science literature, and expresses a situation where very little prior information is available about the location of the Pareto optimal solution set. The initial sample is subsequently iteratively improved using the various algorithmic steps in the employed evolutionary algorithm. It is however possible to significantly improve the efficiency and robustness of evolutionary search for case studies (2) and (3) reported in TRW if we first attempt to create an initial sample that approximates the Pareto tradeoff surface as closely as possible. In our original paper (Vrugt et al., 2003) we suggest such an alternative sampling strategy by first locating the theoretical ends of the Pareto set using classical single objective optimization, and to use traditional first-order statistical theory around the optimal estimators for these individual ob-