基于贝叶斯机器学习的生态模型参数优化方法研究

发布时间：2018-06-20 23:28

本文选题：NUTS + 生态模型　；参考：《地球信息科学学报》2017年10期

【摘要】：参数优化方法是准确估计生态模型参数、降低其不确定性的有效手段。本文提出一种基于贝叶斯机器学习的No-U-Turn Sampler(NUTS)生态模型参数优化方法。NUTS是一种高效的参数优化方法,每次取样中利用递归算法生成候选参数集(二叉树)推断参数的后验信息,如果满足约束条件"非U型回转",不断构建子树更新参数;否则,记录本次抽样的"最优"参数集,并开始下一次取样,直到获取足够样本。该算法在每次取样中充分优化参数,避免因随机游走行为产生冗余抽样,提高了参数优化效率。本文以千烟洲亚热带人工针叶林碳通量模拟为例,基于Pymc3框架利用NUTS参数优化方法实现了碳通量(Net Ecosystem Exchange,NEE)模型参数反演,并与Metropolis-Hastings(MH)方法进行对比。结果表明,本文算法的参数值达到稳定波动时的抽样次数减少了85%左右,参数优化效率提升3倍左右。参数优化后,2种NEE模型中7个参数不确定性降低10%~53%。此外,NEE模拟效果明显提升,模拟值与实测值的R2分别提高23%和17%,RMSE分别降低3%和4%。综上所述,本文提出的参数优化方法对生态领域的参数估计或数据同化工作具有一定的借鉴意义。
[Abstract]:Parameter optimization method is an effective method to estimate the parameters of ecological model accurately and reduce its uncertainty. In this paper, a No-U-Turn Samplern NUTS-based ecological model parameter optimization method based on Bayesian machine learning. NUTS is an efficient parameter optimization method. The recursive algorithm is used to generate the posterior information of the candidate parameter set (binary tree) in every sampling. If the constraint "non-U rotation" is satisfied, subtree update parameters are constantly constructed; otherwise, the "optimal" parameter set of this sampling is recorded and the next sampling begins until sufficient samples are obtained. The algorithm optimizes parameters in every sampling, avoids redundant sampling due to random walk behavior, and improves the efficiency of parameter optimization. In this paper, the numerical simulation of carbon flux of artificial coniferous forest in Qianyanzhou subtropics is taken as an example. Based on Pymc3 framework, the parameter inversion of net Ecosystem Exchange nee) model is realized by using Nuts parameter optimization method, and compared with Metropolis-HastingsMH method. The results show that the sampling times of the algorithm are reduced by about 85% and the efficiency of parameter optimization is increased by about 3 times when the parameters of the algorithm reach stable fluctuation. After parameter optimization, the uncertainty of 7 parameters in the two kinds of NEE models is reduced by 10% and 53%. In addition, the simulation effect of nee was significantly improved, the R2 of simulated value and measured value were increased by 23% and 17%, respectively, and RMSE decreased by 3% and 4%, respectively. To sum up, the parameter optimization method proposed in this paper can be used for reference in the field of ecological parameter estimation or data assimilation.
【作者单位】：沈阳农业大学;中国科学院地理科学与资源研究所生态系统网络观测与模拟重点实验室;中国科学院大学;
【基金】：国家重点研发计划(2016YFC0500204) 国家自然科学基金项目(31501217、41571424) 辽宁省科学技术计划项目(2014201001)
【分类号】：O212;TP18

【相似文献】