IMPRECISE PROBABILITY

IN RISK ANALYSIS

D. Dubois
IRIT-CNRS, Université Paul Sabatier
31062 TOULOUSE FRANCE
 Outline
1. Variability vs incomplete information
2. Blending set-valued and probabilistic
representations : uncertainty theories
3. Possibility theory in the landscape
4. A methodology for risk-informed decision-
making
5. Some applications
 Origins of uncertainty
• The variability of observed repeatable natural
phenomena : « randomness ».
– Coins, dice…: what about the outcome of the next
throw?
• The lack of information: incompleteness
– because of information is often lacking, knowledge
about issues of interest is generally not perfect.
• Conflicting testimonies or reports:inconsistency
– The more sources, the more likely the inconsistency
 Example
• Variability: daily quantity of rain in Toulouse
– May change every day
– It can be estimated through statistical observed data.
– Beliefs or prediction based on this data
• Incomplete information : Birth date of Brazil President
– It is not a variable: it is a constant!
– You can get the correct info somewhere, but it is not available.
– Most people may have a rough idea (an interval), a few know
precisely, some have no idea: information is subjective.
– Statistics on birth dates of other presidents do not help much.
• Inconsistent information : several sources of information
conflict concerning the birth date (a book, a friend, a
website).
 The roles of probability
Probability theory is generally used for
representing two aspects:
1. Variability: capturing (beliefs induced by)
variability through repeated observations.
2. Incompleteness (info gaps): directly modeling
beliefs via betting behavior observation.
These two situations are not mutually exclusive.
Using a single probability distribution to represent
incomplete information is not entirely satisfactory:
The betting behavior setting of Bayesian subjective
probability enforces a representation of partial ignorance
based on single probability distributions.
1. Ambiguity : In the absence of information, how can a
uniform distribution tell pure randomness and ignorance
apart ?
2. Instability : A uniform prior on x∈ [a, b] induces a non-
uniform prior on f(x) ∈ [f(a), f(b)] if f is increasing and
non-affine : ignorance generates information???
3. Empirical falsification: When information is missing,
decision-makers do not always choose according to a
single subjective probability (Ellsberg paradox).
 Motivation for going beyond
probability
• Have a language that distinguishes between uncertainty
due to variability from uncertainty due to lack of
knowledge or missing information.
• For describing variability: Probability distributions 
– but information demanding, and paradoxical for ignorance

• For representing incomplete information : Sets

(intervals).
but a very crude representation of uncertainty
• Find representations that allow for both aspects of
uncertainty.
 Set-Valued Representations of
Partial Information
• A piece of incomplete information about an ill-
known quantity x is represented by a pair (x, E)
where E is a set called a disjunctive (epistemic) set,
• E is a subset of mutually exclusive values, one of
which is the real x.
• (x, E) means « all I know is that x ∈ E »
– Intervals E = [a, b]: incomplete numerical information
– Classical Logic: incomplete symbolic information
E = Models of a proposition stated as true.
• Such sets are as subjective as probabilities
BOOLEAN POSSIBILITY THEORY

If all you know is that x ∈ E then

- You judge event A possible if it is logically consistent with
what you know : A ∩ E ≠ Ø
A Boolean possibility function : Π(A) = 1, and 0 otherwise
- You believe event A (sure) if it is a logical consequence of
what we already know : E ⊆ A
A certainty (necessity) function : N(A) = 1, and 0 otherwise
- This is a simple modal epistemic logic (KD45)
N(A) = 1 - Π(Ac) ≤ Π(A)
Π(A ∪ B) = max(Π(A), Π(B)); N(A ∩ B) = min(N(A), N(B)).
 WHY TWO SET-FUNCTIONS ?

• Encoding 3 extreme epistemic states….

– Certainty of truth : N(A) = 1 (hence Π(A) = 1)
– Certainty of falsity: Π(A) = 0 (hence N(A) = 0)
– Ignorance : Π(A) = 1, N(A) = 0
….. requires 2 Boolean variables!
The Boolean counterpart of a subjective probability
– With one function you can only say believe A or
believe not-A.
but this representation is poorly expressive (no
gradation in uncertanty
 Find an extended representation of
uncertainty
• Explicitly allowing for missing information (= that
uses sets)
• More informative than pure intervals or classical
logic,
• Less demanding and more expressive than single
probability distributions
• Allows for addressing the issues dealt with by both
standard probability, and logics for reasoning
about knowledge.
 Blending intervals and probability

• Representations that may account for variability,

incomplete information, and belief must combine
probability and epistemic sets.
– Sets of probabilities : imprecise probability theory
– Random(ised) sets : Dempster-Shafer theory
– Fuzzy sets: numerical possibility theory
• Relaxing the probability axioms :
– Each event has a degree of certainty and a degree of
plausibility, instead of a single degree of probability
– When plausibility = certainty, it yields probability
 A GENERAL SETTING FOR REPRESENTING
GRADED CERTAINTY AND PLAUSIBILITY
• 2 set-functions Pl and Cr, with values in [0, 1],
generalizing probability, possibility and necessity.
• Conventions :
– Pl(A) = 0 "impossible" ;
– Cr(A) = 1 "certain"
– Pl(A) =1 ; Cr(A) = 0 "ignorance" (no information)
– Pl(A) - Cr(A) quantifies ignorance about A
• Postulates
– If A⊆ B then Cr(A) ≤ Cr(B) and Pl(A) ≤ Pl(B)
– Cr(A) ≤ Pl(A) "certain implies plausible"
– Pl(A) = 1 − Cr(Ac) duality certain/plausible
 Imprecise probability theory
• A state of information is represented by a family P
of probability distributions over a set X.
– For instance an imprecise probabilistic model.
• To each event A is attached a probability interval
[P*(A), P*(A)] such that
– P*(A) = inf{P(A), P∈ P}
– P*(A) = sup{P(A), P∈ P} = 1 – P*(Ac)
• cP = {P, P(A) ≥ P* (A) for all A} is convex
• Usually cP is strictly contained in {P, P ≥ P*}
Random sets and evidence theory
• A family F of « focal » (disjunctive) non-empty
sets  representing
– A collection of incomplete observations (imprecise
statistics).
– Unreliable testimonies
• A positive weighting of focal sets (a random set) :
∑ m(E) = 1 (mass function)
E∈F

• It is a randomized epistemic state where

– m(E) = probability(E is the correct information)
= probability(only knowing”(x, E)")
 Theory of evidence
• degree of certainty (belief) :
– Bel(A) = ∑ m(Ei)
Ei ⊆ A, Ei ≠ Ø
– total mass of information implying the occurrence of A
– (probability of provability)
• degree of plausibility :
– Pl(A) = ∑ m(Ei) = 1 − Bel(Ac) ≥ Bel(A)
Ei ∩ A ≠ Ø
– total mass of information consistent with A
– (probability of consistency)
 Possibility Theory
(Shackle, 1961, Lewis, 1973, Zadeh, 1978)

• A piece of incomplete information "x ∈ E"

admits of degrees of possibility.
• E is mathematically a (normalized) fuzzy set.
• µE(s) = Possibility(x = s) = πx(s)
• Conventions:
∀s, πx(s) is the degree of plausibility of x = s
πx(s) = 0 iff x = s is impossible, totally surprising
πx(s) = 1 iff x = s is normal, fully plausible, unsurprising
(but no certainty)
 POSSIBILITY AND NECESSITY
OF AN EVENT

How confident are we that x ∈ A ⊂ S ? (an event A occurs)

given a possibility distribution π for x on S
• Π(A) = maxs∈A π(s) :
to what extent A is consistent with π
(= some x ∈ A is possible)
The degree of possibility that x ∈ A
• N(A) = 1 – Π(Ac) = min s∉A 1 – π(s):
to what extent no element outside A is possible
= to what extent π implies A
The degree of certainty (necessity) that x ∈ A
 Basic properties

Π(A ∪ B) = max(Π(A), Π(B));

N(A ∩ B) = min(N(A), N(B)).
Mind that most of the time :
Π(A ∩ B) < min(Π(A), Π(B));
N(A ∪ B) > max(N(A), N(B)
Example: Total ignorance on A and B = Ac

Corollary N(A) > 0 ⇒ Π(A) = 1

 POSSIBILITY AS UPPER PROBABILITY

• Given a numerical possibility distribution π, define

P(π) = {P | P(A) ≤ Π(A) for all A}
– Then, Π and N can be recovered
• Π(A) = sup {P(A) | P ∈ P(π)};
• N(A) = inf {P(A) | P ∈ P(π)}
– So π is a faithful representation of a special family of probability
measures
• Likewise for belief functions : P(π) = {P | P(A) ≤ Pl(A), ∀ A}
• Possibility theory corresponds to consonant belief functions
– Nested focal sets: m(E) > 0 and m(F) > 0 imply F⊆E or E ⊆ F
– If and only if Pl(A) = Π(A) and Bel(A) = N(A).
How to build possibility distributions
(not related to linguistic fuzzy sets!!!)

• Nested random sets (= consonant belief functions)

• Likelihood functions (in the absence of priors).
• Probabilistic inequalities (Chebyshev…)
• Confidence intervals (moving the confidence level
between 0 and 1)
• The cumulative PDF of P is a possibility
distribution (accounting for all probabilities
stochastically dominated by P)
 LANDSCAPE OF UNCERTAINTY
THEORIES
BAYESIAN/STATISTICAL PROBABILITY
Randomized points

UPPER-LOWER PROBABILITIES
Disjunctive sets of probabilities

DEMPSTER UPPER-LOWER PROBABILITIES

SHAFER-SMETS BELIEF FUNCTIONS
Random disjunctive sets

Quantitative Possibility theory Classical logic

Fuzzy (nested disjunctive) sets Disjunctive sets
 A risk analysis methodology
1. Information collection and representation
2. Propagation of uncertainty through a
mathematical model
3. Extraction of useful information
4. Decision step
Risk analysis : Information collection step
• Faithfulness principle : choose the type of representation
in agreement with the quantity of available information
– Remain faithful to available information, including information
gaps.
• Simple representations (possibility, generalized p-boxes)
naturally capture expert interval information with
confidence levels, quantiles, means, mode, etc.
– If variability and enough statistical information: probability
distributions.
– If incomplete information on some value : interval, possibility
distribution (fuzzy interval)…
– If parameterized model with ill-known parameters : p-box
• An elicitation procedure to query an expert on available
information
A possibility distribution can be obtained from any
family of nested confidence sets :
P(Αα) ≥ 1 − α, α ∈ (0, 1]

α
1

π
πα
0

FUZZY INTERVAL: N(πα) = 1 − α

 Probability boxes
• A set P = {P: F* ≥ P ≥ F*} induced by two
cumulative disribution functions is called a
probability box (p-box),
• A p-box is a special random interval whose
upper and bounds induce the same ordering.

1
F*
α
F*
0
Eα
 No: I consider it as variable
Yes I wish to represent this
parameter by a unique value ?

I know this value ?

I have a sufficient number of

I have statistical measures ? precise measurements ?

I know an interval that contains it No: I must rely on

with certainty ? Qualitative approach expert judgement The variability is spatial ?
(multi -factor … )

No: it is temporal
I know the support of the I have a sufficient number of
I can provide an estimate of its plausible distribution ? imprecise measurements ?
value ?

I know the type of distribution and

I can provide an interval of more can provide intervals for its Random set
plausible values ? parameters ?

Unique probability
Parametric probability dsitribution
I know that the distribution is family / p-boxes
Interval [a, b] symmetrical

Geostatistical model

Trapezo ïdal fuzzy

interval
I can express
preferences in terms of Legend
intervals ? I can provide its mode
yes

No
Triangular fuzzy
interval Choice criteria

Representation tool

Precise value
 How useful are these practical
representations:
• Cutting complexity:
– Convex sets of probability are very complex
representations
– Random sets are potentially exponential
– P-boxes, possibility distributions and other extensions
are linear, but still encode convex probability set, often
random sets.
• Enriching the standard probability analysis
with meta-information and capabilities for
reasoning about knowledge in the risk analysis
process, while remaining tractable on modern
computers.
 Information propagation step
– Joint Monte-Carlo and interval analysis to be carried
out in the encompassing setting of random sets, with
various independence assumptions.
– Distinction between epistemic (in)dependence and
stochastic independence
• Dependent sources and independent variables
• Independent sources and variables
• No assumption of independence (more difficult to compute)
– Simple representations cannot be preserved via
propagation : general random sets are obtained.
 Hybrid possibility-probability
propagation
• Formal problem: Given a numerical function
f(x1,… xm, y1, … yn),
– assume x1,… xm are independent random variables
– assume y1, … yn are non-interactive possibilistic
variables modelled by fuzzy intervals F1, … Fn

• Then f(x1,… xm, y1, … yn) is a fuzzy random

variable
 Hybrid possibility-probability
propagation
• Computation
– Find N samples a1,… am of x1,… xm using a Monte-
Carlo method.
– For each sample, compute f(a1,… am, F1, … Fn) using
fuzzy interval computation.

• As the output, we get a random fuzzy interval

{(C1,v1) … (CN,vN)} where the Ci are fuzzy
intervals and vi are frequencies
 Presentation of results:
how to interpret results?
• summarize too complex information : the kind of
summary depends on the question of interest:
– P-boxes can address questions about threshold
violations (xout ≥ a ??)
– questions of the form a ≤ xout≤ b?are better addressed
by possibility distributions or generalized p-boxes
– statistical measures of trend, imprecision or variability
• Aim: Lay bare the resulting information gap and
the resulting knowledge on the variability on the
quantity of interest.
Upper and lower distributions of
random fuzzy outputs
Upper and lower distributions of
random fuzzy outputs
 Exploiting Random Fuzzy Intervals
Given a random fuzzy output {(C1,v1) … (CN,vN)} :
– Average imprecision: compute the fuzzy average C =
∑i Civi . The average imprecision is the area under C.
– Observable Variability: defuzzify the Ci’s (midpoint of
the mean interval) and compute the standard deviation
of these numbers
– Potential Variability: Compute the range of the
empirical variance induced by the fuzzy intervals.
– Minimal and maximal average variability: compute
interval variance of the random set with upper
probability ∑I= 1, N νi Πi(A).
 Example (D. Guyonnet, BRGM)
• Generic health risk calculation for the case of the
exposure of persons to a chlorinated organic
solvent (1,1,2-Trichloroethane) via the
consumption of contaminated drinking water.
• The chronic carcinogenic toxicological reference
value for this substance is a unit excess risk (UER),
namely, a probability of excess cancer per unit
daily dose of exposure.
• For a person exposed, we calculate an excess risk
(ER) that is a function of the dose D absorbed by
this person and of the unit excess risk.
 Example (D. Guyonnet, BRGM)
• ER = D · UER
• UER = I·C·FE·DE/BW·LE
• where:
• D = dose of exposure (mg pollutant absorbed, per Kg body
weight and per day),
• I = quantity of water ingested per day (L/d),
• C = concentration of 1,1,2-trichloroethane in the drinking
water (mg/L),
• FE = exposure frequency (d/yr),
• DE = duration of exposure (yr),
• BW = body weight (Kg),
• LE = life expectancy (d).
 Parameter values used for the illustration

Parameter Unit Mode of Lower Mode or Upper limit

representation limit core
Concentration in water µg/L Probability 5 10 20
Ingestion L/d Fuzzy interval 1 1.5 2.5
Exposure frequency d/year Fuzzy interval 200 250 350
Exposure duration Years Probability 10 30 50
Oral slope factor (mg/Kg/d)-1 Fuzzy interval 2 x 10-2 5.7 x 10-2 10-1
 Results (Dubois and Guyonnet, 2010)
Probability (Risk < Threshold) 1.0
0.9
0.8
0.7
0.6
0.5
0.4
Plausibility
0.3
0.2 Belief
0.1 Monte Carlo
0.0
0E+0 1E-5 2E-5 3E-5
Risk threshold
 A decision strategy
The decision is made by evaluating the probability of
a risky event of the form f(x1,… xm, y1, … yn) ≥θ,
– If the lower probability (or expectation) of the risky
event is too high: take action to circumvent the risk.
– If the upper probability is low enough : situation not
risky despite information gap.
– If the probability of the risky event is ill-known
• Collect new information to reduce information gap and run the
analysis again
• If information collection is impossible: build up the most
reasonable confidence index by consensus from the best
experts (possibly being Bayesian again!)
 Decision with imprecise probability:
3 attitudes
• Accept incomparability when comparing imprecise utility
evaluations of decisions.
– Pareto optimality : decisions that dominate other choices for all
probability functions
– E-admissibility : decisions that dominate other choices for at least
one probability function (Walley, etc…)
• Select a single probability measure for each decision.
– Compare lower expectations of decisions (Gilboa)
– Generalize Hurwicz criterion to focal sets with degree of optimism
(Jaffray)
• Select a single probability measure and use expected utility
– Shapley value = pignistic transformation(SMETS)
– By picking a probability measure that achieves a compromise
between pessimistic and optimistic attitudes
 Results (Dubois and Guyonnet, 2010)
Probability (Risk < Threshold) 1.0
0.9
0.8
0.7
0.6
0.5
0.4 Plausibility
0.3 Belief
0.2 Confidence Index
0.1 ? = 1/3
0.0
0E+0 1E-5 2E-5 3E-5
Risk threshold
 Some Applications
• Child exposure to lead on an ironworks
brownfield (Baudrit et al, Consoil, 2005)
• Contamination of groundwater (Baudrit et al. J.
Cont. Hydrology, 2007)
• Radioactivity of cowmilk near la Hague (Baudrit
& Chojnacki
• synthesis of multiple sources of information
applied to nuclear computer codes (Destercke and
Chojnacki, Nuclear Eng.& Design, 2008)
• Underground CO2 confinement (ANR project
CRISCO2)
 Diana example
• Compute the probability
– that a mission fails
– that a plane component fails
• From
– statistical knowledge coming from databases reporting
previous incidents
– Diagnosis studies (fault trees, etc.)
– Prognosis studies (duration predictions)
• In order to help in operability and maintenance
decisions
 Main tools
• Computing probability of faults from the
knowledge of fault trees computed by
Altarica software.
• Compute probabilities of risky events of
interest using algorithms exploiting Markov
chain models of component behavior or
missions
• But uncertainty about the probabilistic data
 Modelling uncertainties for DIANA

• Probabilities are ill-known : what is the impact

of this imprecision on the results and the decision
process?
• Computing probabilities attached to cut-sets
defining disjunctions of causes of a risky event
(Boolean formulas), from knowledge of
probabilities of atomic events
• Study of interval-valued Markov chains
– DO EFFICIENT ALGORITHMS IN THE
COMPLETE INFORMATION CASE STILL APPLY ?
 The quality of probabilistic information

• Where do basic probabilities come from ?

• Modelling imprecise probabilities : intervals,
fuzzy (confidence) intervals ???
• Extreme probabilities : approximate
calculations are enough ?
• Propagate imprecision from data to
probabilities of events of interest.
• Question independence assumptions ?
 Conclusion
• There exists a coherent range of uncertainty
theories combining interval and probability
representations.
– Imprecise probability is the proper theoretical umbrella
– The choice between subtheories depends on how
expressive it is necessary to be in a given application.
– There exists simple practical representations of
imprecise probability
• Many open problems, theoretical, and
computational, remain.
• How to get this general non-dogmatic approach to
uncertainty accepted by traditional statisticians?