人工智能培训

搜索

人工智能论文:贝叶斯代理学习基于动态模拟器的回归问题(Bayesian surrogate learning in dynamic simulator-base

[复制链接]
laishi 发表于 2019-1-28 11:50:36 | 显示全部楼层 |阅读模式
laishi 2019-1-28 11:50:36 206 0 显示全部楼层
人工智能论文:贝叶斯代理学习基于动态模拟器的回归问题(Bayesian surrogate learning in dynamic simulator-based regression  problems)表征物理系统的参数(或隐藏变量,控制变量)的未知值的估计通常依赖于测量数据与由系统的一些数值模拟器产生的合成数据的比较,因为参数值是变化的。这个过程经常遇到两个主要的困难:如果系统模型复杂,为每个考虑的参数值生成合成数据可能在计算上是昂贵的;并且参数空间的探索可能是低效的和/或不完整的,典型的例子是当探测陷入对象函数的局部最优时,其表征测量数据和合成数据之间的不匹配。提出了解决这两个问题的方法,其中:使用深度重复网络(DRN)构建模拟计算上昂贵的系统模拟器的代理模型(或代理);并且采用嵌套采样(NS)算法来对参数空间进行有效且稳健的探索。该分析在贝叶斯环境中进行,其中样本表征参数的全关联后验分布,从中容易导出参数估计和不确定性。在一些数值例子中将所提出的方法与传统方法进行比较,结果证明可以将参数估计过程加速至少一个数量级。
The estimation of unknown values of parameters (or hidden variables, controlvariables) that characterise a physical system often relies on the comparisonof measured data with synthetic data produced by some numerical simulator ofthe system as the parameter values are varied.This process often encounterstwo major difficulties: the generation of synthetic data for each consideredset of parameter values can be computationally expensive if the system model iscomplicated;and the exploration of the parameter space can be inefficientand/or incomplete, a typical example being when the exploration becomes trappedin a local optimum of the objection function that characterises the mismatchbetween the measured and synthetic data.A method to address both these issuesis presented, whereby: a surrogate model (or proxy), which emulates thecomputationally expensive system simulator, is constructed using deep recurrentnetworks (DRN);and a nested sampling (NS) algorithm is employed to performefficient and robust exploration of the parameter space.The analysis isperformed in a Bayesian context, in which the samples characterise the fulljoint posterior distribution of the parameters, from which parameter estimatesand uncertainties are easily derived.The proposed approach is compared withconventional methods in some numerical examples, for which the resultsdemonstrate that one can accelerate the parameter estimation process by atleast an order of magnitude.人工智能论文:贝叶斯代理学习基于动态模拟器的回归问题(Bayesian surrogate learning in dynamic simulator-based regression  problems) LC5UapWApaxPp0o0.jpg
URL地址:https://arxiv.org/abs/1901.08898     ----pdf下载地址:https://arxiv.org/pdf/1901.08898    ----人工智能论文:贝叶斯代理学习基于动态模拟器的回归问题(Bayesian surrogate learning in dynamic simulator-based regression  problems)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

laishi当前离线
新手上路

查看:206 | 回复:0

快速回复 返回顶部 返回列表