validation metrics [-1.5cm] - ncsu
TRANSCRIPT
Validation MetricsKathryn Maupin
Laura Swiler
June 28, 2017
Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia LLC, a wholly ownedsubsidiary of Honeywell International Inc. for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525. SAND NO.
2017-XXX
Overview
Model Validation
Data Classification
Validation Metrics
Examples
Conclusions
June 28, 2017 2
Model Validation
The comparison of experimental observations with model output
Observed values contain uncertaintiesExperimental measurement errorLimited/incomplete dataModel form errorParameter uncertaintyApproximation/discretization error
Validation Metric: quantifies the difference between physical andsimulation observations
Observations may be considered with or without uncertainties
June 28, 2017 3
Data Types
Type 1: Experimental and model values are treated withoutuncertainty
Type 2: Experimental values are treated as uncertainNominal value, given by expertsStandard deviation, calculated using multiple experiments
Type 3: Experimental and model values are treated as uncertainUncertainty analysis
June 28, 2017 4
Data TypesType 1 (no uncertainties)
Type 2 (experimental uncertainty)
June 28, 2017 5
Data TypesType 1 (no uncertainties)
Type 2 (experimental uncertainty)
June 28, 2017 5
Data Types
Type 3 (experimental and model uncertainty)
June 28, 2017 6
Classification of Validation MetricsMetric Type 1 Type 2 Type 3Root Mean Square ✓Minkowski Distance ✓Simple Cross Correlation ✓Normalized Cross Correlation ✓Normalized Zero-Mean Sum of Squared Distances ✓Moravec Correlation ✓Index of Agreement ✓Sprague-Geers Metric ✓Normalized Euclidean Metric ✓Mahalanobis Distance ✓Hellinger Metric ✓ ✓Kolmogorov-Smirnoff Test ✓ ✓Kullback-Leibler Divergence ✓Symmetrized Divergence ✓Jensen-Shannon Divergence ✓Total Variation Distance ✓
June 28, 2017 7
Example Validation Metrics
Type 1 DataMinkowski Distance (lp Distance)
d =
(∑i|Pi − Di|p
)p
Type 2 DataMahalanobis Distance
d =
√(P − D)T Σ−1
D (P − D)
Type 3 DataKullback-Leibler Divergence
DKL(ND∥NP) =1
2
[tr(Σ−1
P ΣD) + (P − D)TΣ−1P (P − D)− k + ln
(detΣPdetΣD
)]Kolmogorov-Smirnov Test
DKS = sup |FP(x)− FD(x)|
June 28, 2017 8
Example Validation Metrics
Type 1 DataMinkowski Distance (lp Distance)
d =
(∑i|Pi − Di|p
)p
Type 2 DataMahalanobis Distance
d =
√(P − D)T Σ−1
D (P − D)
Type 3 DataKullback-Leibler Divergence
DKL(ND∥NP) =1
2
[tr(Σ−1
P ΣD) + (P − D)TΣ−1P (P − D)− k + ln
(detΣPdetΣD
)]Kolmogorov-Smirnov Test
DKS = sup |FP(x)− FD(x)|
June 28, 2017 9
Example Validation Metrics
Type 1 DataMinkowski Distance (lp Distance)
d =
(∑i|Pi − Di|p
)p
Type 2 DataMahalanobis Distance
d =
√(P − D)T Σ−1
D (P − D)
Type 3 DataKullback-Leibler Divergence
DKL(ND∥NP) =1
2
[tr(Σ−1
P ΣD) + (P − D)TΣ−1P (P − D)− k + ln
(detΣPdetΣD
)]Kolmogorov-Smirnov Test
DKS = sup |FP(x)− FD(x)|
June 28, 2017 10
Example 1
“Experiment”f(x) = 1.1 log(10x) + ε
Modelg(x) = log(10x)
withε = measurement errorxi = 1, 2, . . . 20
Type 1 Data
Metric Value Min MaxRoot Mean Square 4.4706× 10−1 0 ∞Average Relative Minkowski Distance
p = 1 8.9337× 10−2 0 ∞p = 2 3.0483× 10−3 0 ∞
June 28, 2017 11
Example 1Type 2 Data
June 28, 2017 12
Example 1Type 2 Data
Metric Value (5%) Value (10%) Min MaxAverage Mahalanobis Distance 3.5891 1.7946 0 ∞
June 28, 2017 13
Example 1Type 3 Data
June 28, 2017 14
Example 1Type 3 Data
Metric Value (5%) Value (10%) Min MaxKullback-Leibler Divergence
Total 1.5608× 102 3.9162× 101 0 ∞Average per point 7.8041 1.9581 0 ∞
June 28, 2017 15
Example 1Type 3 Data
Metric Value (5%) Value (10%) Min MaxKolmogorov-Smirnov Test
Maximum 9.7072× 10−1 7.2466× 10−1 0 1Average 9.3431× 10−1 6.4907× 10−1 0 1
June 28, 2017 16
Example 2
“Experiment”f(x) = sin2(x) + 5 + ε
Modelg(x) = sin2(x) + θ
withε = measurement error
Type 1 DataMetric Value Min MaxRoot Mean Square 1.0242× 10−1 0 ∞Average Relative Minkowski Distance
p = 1 1.5291× 10−2 0 ∞p = 2 1.8371× 10−2 0 ∞
June 28, 2017 17
Example 2Type 2 Data
Metric Value (5%) Value (10%) Min MaxAverage Mahalanobis Distance 7.3463× 10−1 3.6817× 10−1 0 ∞
June 28, 2017 18
Example 1Type 3 Data
Metric Value (5%) Value (10%) Min MaxKullback-Leibler Divergence
Total 5.8205× 101 2.4860× 102 0 ∞Average per point 4.0703× 10−1 1.7384 0 ∞
June 28, 2017 19
Example 1Type 3 Data
Metric Value (5%) Value (10%) Min MaxKolmogorov-Smirnov Test
Maximum 7.2276× 10−1 5.8143× 10−1 0 1Average 2.6065× 10−1 3.0401× 10−1 0 1
June 28, 2017 20
Conclusions
Computational models require validation before they can bereliably used in prediction scenarios
Metrics exist for deterministic data and probabilistic (uncertain)data
Choice of metric is application and quantity of interest dependent
Future work: develop a guide to determinechoice of validation metricvalidation tolerance
June 28, 2017 21