Comment on “High‐dimensional posterior exploration of hydrologic models using multiple‐try DREAM (ZS) and high‐performance computing” by Eric Laloy and Jasper A. Vrugt
Comment on “High‐dimensional posterior exploration of hydrologic models using multiple‐try DREAM (ZS) and high‐performance computing” by Eric Laloy and Jasper A. Vrugt
复制标题
DOI：
10.1002/2012wr013341
复制
发表时间：
2014-03
影响因子：
5.4
通讯作者：
Wei Chu;Tiantian Yang;Xiaogang Gao
中科院分区：
文献类型：
--
作者：
Wei Chu;Tiantian Yang;Xiaogang Gao
Laloy和Vrugt[2012年](DOI：10.1029/2011WR010608，以下简称LV)最近引入了一种新的马尔可夫链蒙特卡罗算法，名为MT-DREAM(ZS)。为了说明这种新算法的性能改进，LV将其与另一种自适应MCMC方法DREAM(ZS)[Vrugt等人，2009]和直接搜索方法SP-UCI[Chu等人，2010]进行了比较。用作基准的模拟案例是萨克拉门托土壤水分核算(SAC-SMA)模型和基廷等人开发的包含241个参数的地下水模型。[2010]。根据他们的实验结果，LV得出结论，SP-UCI在这两种模型上的表现都很差，因此LV提出了SP-UCI处理复杂和多模式搜索空间的能力的问题。实验表明，MT-DREAM算法在搜索高维模型的全局最优解方面明显优于DREAM算法(ZS)和SP-UCI算法。我们祝贺LV的新算法以及MT-Dream(ZS)和Dream(ZS)包中提供的测试用例的一致性能。然而，在重复了LV的论文中的另外两个结论性实验后，我们想表达我们对SP-UCI对这两个模型的性能和能力的不公平判断的担忧。此外，我们对实验结果的可重复性以及MT-DREAM(ZS)算法在评估LV报告的关于SACSMA模型和241参数地下水模型上的MT-DREAM和SP-UCI算法的复杂后验分布的有效性感到困惑。
Laloy and Vrugt [2012] (doi:10.1029/2011WR010608, hereafter referred to as LV) recently introduced a new Markov Chain Monte Carlo (MCMC) algorithm, entitled MT-DREAM(ZS). To illustrate the improved performance of this new algorithm, LV compared it against another adaptive MCMC approach, DREAM(ZS) [Vrugt et al., 2009], and a direct search method, SP-UCI [Chu et al., 2010]. The modeling cases used as benchmarks were the Sacramento Soil Moisture Accounting (SAC-SMA) model and a 241-parameter groundwater model developed by Keating et al. [2010]. Based on their experimental results, LV concludes that ‘‘SP-UCI performs poorly’’ on both models, therefore LV raises the question about the ability of SP-UCI to deal with complex and multimodal searching spaces. Also, the experiments show that MT-DREAM algorithm demonstrates substantial improvement over DREAM(ZS) and SP-UCI for searching the global optima of high-dimensional models. We congratulate LV on their new algorithm and the consistent performances for the test examples provided in the MT-DREAM(ZS) and DREAM(ZS) packages. However, after repeating other two conclusive experiments in LV’s paper, we would like to communicate our concerns regarding the unfair judgments about the performance and ability of SP-UCI for these two models. Also, we are puzzled about the reproducibility of the experimental results and the effectiveness of the MT-DREAM(ZS) algorithm in assessing complex posterior distributions reported by LV about both the MT-DREAM and SP-UCI algorithms on the SACSMA model and the 241-parameter ground water model.