Reconstructing Input Models in Stochastic Simulation

Reconstructing Input Models in Stochastic Simulation
复制标题

在随机模拟中重建输入模型

DOI:
--
复制
发表时间:
2016
期刊:
影响因子:
--
通讯作者:
Bo Zhang
Bo Zhang
中科院分区:
--
文献类型:
--
作者:
A. Goeva;H. Lam;Huajie Qian;Bo Zhang

文献摘要

被引文献

相似文献

我们调查的问题,非参数化校准的输入模型在随机模拟,只给出了输出数据的可用性。虽然模拟输入不确定性的研究集中在使用输入数据,我们的逆模型校准问题出现在各种情况下,资源和操作的限制,阻碍了直接观测的输入数据。我们提出了一个基于矩的最大熵框架来推断输入模型,通过匹配模拟输出和真实的世界输出之间的统计。为了绕过我们制定的随机约束的方法上的困难,我们提出了一个随机二次罚函数的方法,将问题转化为一个序列的最小二乘问题,其中序列中的每个元素可以通过有效的随机近似算法来解决。我们分析了该方法的统计特性,并通过数值实验证明了其恢复输入模型的能力。
We investigate the problem of nonparametrically calibrating the input model in stochastic simulation, given only the availability of output data. While studies on simulation input uncertainty have focused on the use of input data, our inverse model calibration problem arises in various situations where resource and operational limitations hinder the direct observability of input data. We propose a moment-based, maximum entropy framework to infer the input model, by matching statistics between the simulation output and the real- world output. To bypass the methodological difficulties from the stochastic constraints in our formulation, we propose a stochastic quadratic penalty method that converts the problem into a sequence of least-square problems, where each element in the sequence can be solved by efficient stochastic approximation algorithms. We analyze the statistical properties of our method and demonstrate its ability to recover input models with numerical experiments.