Title: A Comparison of Information Management using Imprecise Probabilities and Precise Bayesian Updating o
1A Comparison of Information Management using
Imprecise Probabilities and Precise Bayesian
Updating of Reliability Estimates
- Jason Matthew Aughenbaugh, Ph.D.
- jason_at_arlut.utexas.edu
- Applied Research Laboratories
- University of Texas at Austin
- Jeffrey W. Herrmann, Ph.D.
- jwh2_at_umd.edu
- Department of Mechanical Engineering and
Institute for Systems Research - University of Maryland
- Third International Workshop on Reliable
Engineering Computing, NSF Workshop on Imprecise
Probability in Engineering Analysis Design,
Savannah, Georgia, February 20-22, 2008.
2Motivation
- Need to estimate reliability of system with
components of uncertain reliability. - Which components should we test to reduce
uncertainty about system reliability?
3Introduction
Existing information
Prior characterization
Is it relevant?
Data
Is it accurate?
Statistical modeling and updating approach
New experiments
Updated / posterior characterization
4Statistical Approaches
- Compare the following approaches
- (Precise) Bayesian
- Robust Bayesian
- sensitivity analysis of prior
- Imprecise probabilities
- actual true probability is imprecise
- the imprecise beta model
Different philosophical motivations, but
equivalent math. for this problem
5Is precise probability sufficient?
- Problem equiprobable
- Know nothing or know they are equally likely?
- Why does it matter?
- Engineer A states that input values 1 and 2 have
equal probabilities - Engineer B is designing a component that is very
sensitive to this input - Should Engineer B proceed with a costly but
versatile design, or study the problem further? - Case 1 Engineer A had no idea, so stated equal.
Study good - Case 2 Engineer A performed substantial
analysis. Additional study wasteful.
6Moving beyond precise probability
- Start with well established principles and
mathematics - Conclude it is insufficient
- Abandon probability completely?
- Relax conditions, extend applicability?
Think sensitivity analysis. How much do
deviations from a precise prior matter?
7Robust Bayes, Imprecise Beta Model
- Instead of one prior, consider many (a set)
Cumulative Probability
?
8Problem Description
- A simple parallel-series system, some info
- Assume we can test 12 more components
- How should these tests be allocated?
- A single test plan can have different outcomes
- Compare different scenarios of existing
information
9Multiple Outcomes of Experiement
- Precise probability
- Consider one outcome test A 12 times, 2 fail
- Get one new posterior precise parameters
- Consider all possible outcomes test A, get
- Get a new posterior for each possible outcome
sets of parameters - Imprecise probability
- One outcome, one SET of posteriors
- Multiple outcomes, SET of SETS of posteriors
How measure uncertainty? How make comparisons
and decisions?
10Metrics of Uncertainty Precise Distributions
- Variance-based sensitivity analysis (SVi)
- (Sobol, 1993 Chan et al., 2000)
- variance of the conditional expectation / total
variance - focuses on status quo, next (local) piece of info
- testing a component with a large sensitivity
analysis should reduce variance of system
reliability estimate - Mean and variance observations
- Posterior variance
11Metrics of Uncertainty Imprecise Distributions
- Imprecise variance-based sensitivity analysis
(Hall, 2006) - Does not worry about outcomes local metric
- Mean and variance dispersion
- Imprecision in the mean
- Imprecision in the variance
12Scenarios with Precise Distributions
Scenario 1 priors
- Components have beta distributions for the prior
distributions of failure probability - Scenario 1
- System failure probabilitymean 0.2201
variance 0.0203 - Scenario 2
- System failure probability mean
0.1691variance 0.0116
X
X
Scenario 2 priors
13Scenario 1 Results
- Variance-based sensitivity analysis
Best worst-case
Best best-case
14Scenario 1 Results
2
1
15Scenario 2 Results
- Variance-based sensitivity analysis
Best worst-case
Best best-case
16Scenario 2 Results
2
1
17Scenario 3 Imprecise Distributions
- Component failure probabilities are modeled using
imprecise beta distributions - System failure probability an imprecise
distribution - Mean 0.2201 to 0.4640
- Variance 0.0136 to 0.0332
- Imprecise variance-based sensitivity analysis
Since failure probability of B is poorly known,
we allow for a range. Scenario 3 comparable to
precise scenario 1.
18Posterior Variance Analysis
Smallest variances, and smallest imprecision in
variances.
19Results for Scenario 3
Sample results12, 0, 0, 0, 12, 0, 6, 6, 0
Convex hull of results 12, 0, 0, 0, 12, 0,
6, 6, 0
Convex hull of results0, 0, 12, 6, 0, 6,
0, 6, 6
Convex hull of results0, 12, 0, 4, 4, 4,
6, 6, 0, 0, 6, 6
20Scenario 4 Imprecise Distributions
- Component failure probabilities are modeled using
imprecise beta distributions - System failure probability is also an imprecise
distribution - Mean 0.1691 to 0.2880
- Variance 0.0100 to 0.0173
- Imprecise variance-based sensitivity analysis
Compared to scenario 3, the failure probability
of C is reduced. This makes it comparable to
precise scenario 2.
21Results for Scenario 4
Convex hull of results0, 0, 12, 6, 0, 6,
0, 6, 6
Convex hull of results12, 0, 0, 0, 12, 0,
6, 6, 0
Convex hull of results12, 0, 0, 4, 4, 4,
0, 6, 6
22Discussion / Future Work
- Multiple sources of uncertainty
- Existing knowledge
- Results of future tests
- How do we prioritize different aspects?
- Variance or imprecision reduction?
- Best case, worst case, average case of results?
- Incorporate economic/utility metrics?
- Other imprecision/total uncertainty measures?
- Breadth of p-boxes (Ferson and Tucker, 2006 )
- Aggregate uncertainty, others(Klir and Smith,
2001)
23Summary
- Shown how to use different statistical approaches
for evaluating experimental test plans - Used direct uncertainty metrics
- Variance-based sensitivity analysis
- Precise and imprecise
- Posterior variance
- Dispersion of the mean and variance
- Imprecision in the mean and variance
24Thank you for your attention.
- Questions? Comments? Discussion?
This work supported in part by the Applied
Research Laboratories at UT-Austin Internal IRD
grant 07-09
25SVi
26Formulae
. The mathematical model for the reliability of
the system shown in Figure 1 follows.