2023-11-20 论文 Uncertainty Quantification in Machine Learning for Engineering Design and Health Prognostics

发布于 2023年11月20日  •  2 分钟  • 516 字

Abstract

• types
  • 第一种分类
    • data uncertainty (measurement noise)
    • model uncertainty ( limited data)
  • 第二种分类
    • epistemic uncertainty
      • 认知上的不确定性，通常是由于没有足够的知识（数据）而产生
      • can be reducible
      • 分为两类
        • model-form uncertainty
          • 由于模型的选择导致，例如architectures, activation functions or kernel functions
        • parameter uncertainty
          • 在训练过程产生，由于数据不够导致
    • aleatory uncertainty
      • stems from physical systems, 具有随机性, cannot be reducible
      • e.g. noises
      • 这种类型的不确定性在ML模型里面被看成是似然函数的一部分(a part of the likelihood function)
      • 也被叫做data uncertainty
      • 捕捉这种不确定性的方式有：同方差 homoscedastic和异方差 heteroscedastic
    • 例子：
      • test data和train data不同分布：epistemic uncertainty (model performs poorer in extrapolation than in interpolation)
      • 测量数据由仪器导致的误差是aleatory Unc， 大试如果由于精度原因导致，则属于epistemic unc，因为提高精度可以减少这个误差
• causes
• methods:
  • Gaussian process regression
    • a ML method with UQ capability
    • 一般不用来quantify uncertainty of a final surrogate
    • 一般用来在高度不确定的采样空间里采样，来减少训练样本的数量
      • to build an accurate surrogate within some lower and upper bounds of input variables
      • to find a globally optimally design for black-box objective function
    • 一般不评估GPR的UQ质量
      • 因为预测一般在pre-defined design bounds
  • Bayesian neural network
    • Monte Carlo dropout as an alternative to traditional Bayesian neural network
  • neural network ensemble
    • neural network ensemble consisting of multiple neural networks
  • deterministic UQ methods
• metrics
  • classification
    • probability can be viewed as uncertainty
  • regression
    • confidence interval :
      • 没看懂： prediction may be 120 ± 15, in weeks, which represents a two-sided 95% confidence interval (i.e.,∼1.96 standard deviations subtracted from or added to the mean estimate assuming the model-predicted RUL follows a Gaussian distribution).