Title: Expert Judgment Techniques and Distribution Sensitivity Analyses in RMPS Methodology
1 Expert Judgment Techniques and Distribution
Sensitivity Analyses in RMPS Methodology
Ricardo Bolado Lavín Nuclear Safety Unit,
Institute for Energy, European Commission
DG-JRC http// www.energyrisks.jrc.nl http//www
.jrc.nl
2RMPS methodology
3 Identification of relevant parameters
- Use of the Analytical Hierarchy Process (AHP). It
allows to create a hierarchy of parameters in the
case studied (rank parameters according to their
importance) - Building a hierarchy to decompose the problem
- Definition of the top goal (top level in the
hierarchy) - Create lower levels searching for factors that
affect the upper level and are affected by others
in the lower level - Locate basic parameters in the lowest part of the
hierarchy - Use of pair wise comparisons
- For each element in each level build a pair wise
comparison matrix to assess importance of
elements in the level just below it. - Normalized eigenvalues of the pair wise matrix
comparison provide the ranking - Index available to detect inconsistent
assessments.
4 Identification of relevant parameters (2)
5Quantification of uncertainties
- The method to be applied depends on the quantity
of data available - Many data available
- Classical estimation methods Chi-square,
Kolmogorov - Not many data available
- Bayesian estimation methods
- Scarce or almost no data
- Expert judgment methods
6Distribution sensitivity analysis
- What happens with our output variable (e.g.
probability of success) if the distributions of
our inputs are changed? - There are several reasons for the experts to
provide more than one input vector distribution - Experts may show reluctance to provide accurate
answers/distributions (under confident experts). - Experts may acknowledge that the actual
distribution could be slightly different from the
one provided by them. - In some cases the experts consider that the
distribution should be divided into different
distributions (conditional distributions). - Sometimes experts just disagree and provide
different distributions - A check of the influence of those different
distributions may be worthy.
7The trivial solution
- Let us assume that we want to estimate the
influence of changing the distribution of one
input parameter - Take a sample of size n under the original
distribution - Propagate the uncertainty through the model
- Estimate the reliability under the original
distribution - Take a sample of size n under the alternative
distribution - Propagate the uncertainty through the model
- Estimate the reliability under the alternative
distribution - Compare both reliabilities.
8The trivial solution (2)
- Let us assume that we want to estimate the
influence of changing the distribution of k input
parameters - Repeat the process described in the first three
points of last slide. - Repeat the process described in steps 4 to 7 in
last slide k times. - Total number of computer code runs n kn
- Let us assume that we want to estimate the
influence of changing simultaneously the
distribution of 2 or more input parameters - The problem becomes a combinatorial problem
(forget the trivial method).
9Available acceptable solutions
- Two methods proposed by Beckman and McKay (1987)
- The re-weighting method (based on the idea of
importance sampling) - Provides unbiased estimators for the mean of the
output variable - The rejection method (based on the idea of
acceptance-rejection sampling) - Estimates in an unbiased way the distribution
function of the output variable (adequate for our
objectives)
10The problem
- Let us consider a k components vector of input
parameters X. - Let us assume that two different distribution
functions f1(x) and f2(x) may be assigned to that
input parameter vector. - Let us also assume that the first one is the
reference one and the second one is the one we
are addressing for sensitivity. - For simplicity, let us consider only one output
variable Y. - Let us call F1(y) and F2(y) to the distribution
functions of Y when X follows the distributions
f1(x) and f2(x) respectively. - The challenge
- To use only one sample to estimate simultaneously
properties of F1(y) and F2(y) in order to compare
them.
11The rejection method
- There are two conditions for the application of
this method - First, as in the previous method, the support of
f2(x), R2, must be contained in the support of
f1(x), R1 and - Second, the quotient f2(x)/f1(x) must be bounded.
Let us assume that the bound is M.
12The rejection method (2) - Implementation
- Step 1? To get a sample of size n of the input
random vector under the reference distribution
X1, X2, , Xn. Their actual values will be x1,
x2, , xn. - Step 2 ? To run the code for those n sets of
inputs in order to get the corresponding sample
of the output variable Y1, Y2,, Yn. Their
actual values will be y1, y2,, yn. - Step 3? For each sample xi, take a sample of the
uniform distribution Vi between 0 and Mf1(xi). - Step 4? To retain in the sample the corresponding
output value Yiyi if the realisation vi of Vi is
less or equal to f2(xi), otherwise reject that
value from the sample. - Step 5? To consider the values of Y remaining in
the sample as a random sample of Y under f2(x).
That sample of size kn will be used to build up
an empirical distribution function that estimates
the actual distribution function of Y under
f2(x). Each step in that empirical distribution
function will have a height 1/k.
13The rejection method (3) - Example
14The rejection method (4) - Example
15Problem Inefficiency
- The efficiency of this method is 1/M (probability
of a random realization of X obtained under f1(x)
being retained as a random realization under
f2(x)). - Intuitive justification the quotient between the
hypervolume under Mf1(x) and the hypervolume
under f2(x) is M, so that, in the average,
according to the acceptance criterion, only a
fraction 1/M of the observations obtained under
f1(x) will remain as a sample under f2(x)
16Proposed solution
17Proposed solution (Extended rejection method)
- Applying the rejection method an infinite number
of times and averaging converges to a solution
(red stepwise line in previous slide) that uses
all the information in an optimum way - The distribution obtained is a discrete random
variable with probability - for each sample value Y(xi)
- The estimator for the mean
- under the new distribution is
18Application of the method
- RP2_B case
- Success criterion Pressure comes below 40 bar
before the end of the scenario - Possible variations in input distributions
- Mean varies up to 25 of the standard deviation
- Standard deviation varies up to 10
19Application of the method
20Application of the method
21Conclusions
- There are methods available for phenomena and
parameter identification and screening - There are methods available for uncertainty
quantification - There are methods available for assessing
(performance) reliability - There are methods for computing distribution
sensitivity - Are there systems available to apply these
methods on them? -
- Fortunately it seems so!!!