Download DYNAMIC STRATEGIC PLANNING - Massachusetts Institute of

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Choice modelling wikipedia , lookup

Data assimilation wikipedia , lookup

Transcript
Uncertainty Assessment

The quantified description of the uncertainty
concerning situations and outcomes
Presentation Objectives:
1. Concepts

– The problem and issues
– Means of assessment
2. Analytic Procedures
– Regression analysis
– Bayes Theorem and Likelihood Ratios
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 1 of 35
Part 1
Concepts

The problem and issues

Means of assessment
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 2 of 35
The Problem


New Design Paradigm requires us to
determine Distribution of Uncertainties
How can we do this?
Issues:
1. Procedure depends on data available

– No one way
2. Systematic Biases exist throughout
– Need to be aware and careful
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 3 of 35
Methods of Assessment -- Types

Logic
– Example: Probability (Queen) in a deck of cards
– Primary Content of Intro. Probability Subjects
– Not really useful to assess System Uncertainty



Frequency
Statistical Methods
Judgment
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
More on each
coming up
Richard de Neufville ©
Uncertainty assessment
Slide 4 of 35
Frequency Method

Analysis Approach
– Assemble available data
– Find frequency of occurrence of event of interest


Example 1:
Probability of Rainfall in a location
– Use historical data from rain gauges

Example 2:
P (failure of major dams) ~ 0.00001/dam/year
– Source: Baecher et al, data through 1960
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 5 of 35
Frequency Method -- Issues

Assumes process is “stationary”, that is
– Future will be like past ; no change in mechanism
– Is this always true?
– No! Global warming may change rainfall pattern

Assumes data representative
– That it represents item of interest
– When might this not be true?
– Technology of construction changes…
Different types of dams, etc.
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 6 of 35
Statistical Models

Analysis Approach
– Assemble Data, Choose a set to analyze
– Posit a set of variables (X = price, etc) of interest
– Posit form of variables (Price, ΔPrice, Relative P…)
– Posit form of equations linking them, f(X)
– Do statistical analysis (example: least squares)


Example:
Future Demand = f(price, income, etc.) + error
– Obtain: D = a0 + a1Pb + a2Pc +…
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 7 of 35
Statistical Models – General Issues

All the issues associated with Frequency
Method
– Statistical Analysis is a more sophisticated
version of Frequency Method



Specifically, assumes that
Process is stationary
Data are representative
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 8 of 35
Statistical Models – Data Issues


Looks precise and technical, BUT
GIGO -- “Garbage in, Garbage out”
– Results depend closely on data set chosen
(see slides in regression analysis later on0
– Consider history of oil prices: which period
should we choose?

Everything available?

Since OPEC 1, 1974?

Last 20 years?

Last 10 years?
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 9 of 35
History of Oil Prices
Average Annual Crude Oil Price
$120.00
Price (USD)
$100.00
$80.00
$60.00
$40.00
$20.00
19
46
19
50
19
54
19
58
19
62
19
66
19
70
19
74
19
78
19
82
19
86
19
90
19
94
19
98
20
02
20
06
$0.00
Years
Nominal
Inflation Adjusted 2007
Source: http://www.inflationdata.com/inflation/Inflation_Rate/Historical_Oil_Prices_Table.asp
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 10 of 35
Statistical Models – Data Issues


Conceptual “house of cards”
Consider Assumptions
– Posit a set of variables
– Which ones? Note: adding any factor increases
statistical fit
– Posit form of variables (Price, ΔPrice, etc.)
– Which form? When you choose travel, what
influences you: total, relative or differential cost?
– Posit form of equations linking them, f(X)
– Which form? Linear, exponential, log…
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 11 of 35
Statistical Models – Time Issues

Special issue when model seeks to
determine “rate of change over time”
Y = a ert
r = rate per period, t

Data is a “time series”, as for oil price


Issue: Any variable that grows or decreases
reasonably also fits the data well
– Example: Los Angeles Air Travel correlates well
with “prisoners in Oregon”, “divorces in France”
– Another possible form of GIGO

Easy to get good correlation  Causality
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 12 of 35
Expert Judgment Method
• Also known as “Subjective Probability”
• Analysis Method
– Identify “experts”
– Ask them questions such as
“What will be performance of a new technology?

Several variations, concerning how group of
experts interact, come to consensus or not…
– “Delphi” method, named after mythical oracle
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 13 of 35
Expert Judgment -- Issues
• Who are the experts?
– May be easy to identify who is not expert
– BUT, knowledgeable insiders may have
“trained incapacity” – be blind to new factors

Overconfidence
– Distribution typically much broader than
we imagine (remember class experience)

Insensitivity to New Information
– Information typically should cause us to
change opinions more than it does
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 14 of 35
Summary of Part 1





Available methods each have difficulties.
These are conceptual – will not be eliminated
by better math or technique!
Such issues contribute to the fact that
“forecasts are ‘always’ wrong
We have to be properly modest about how
far analysis will take us.
Do the analysis, but appreciate weaknesses
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 15 of 35
Part 2
Analytic Procedures



Regression analysis
Bayes’ Theorem
… and Likelihood Ratios
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 16 of 35
Regression Analysis -- Concept





Object: “best fit” of data to function Y = f(X)
Y often called the “dependent” variable
Xs then called the “independent” variables
Xs then “explain” the variation in Y
Think of Y as deformation of spring under
load, X, …
– Changes in X do not account for all deformation
observed – other factors and errors
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 17 of 35
Regression Analysis – Set up

Object: “best fit” of data to function Y = f(X)

Assumptions about data
– Each measurement of Independent variables (ex:
Weights, Xi ) are correct
– Dependent variables (ex: Deflection, Yi ) have ‘errors’

Model then is: Deflection = a (Weight)
– And we recognize that measured values Ymi will
differ from predicted values Yi will by an ‘error’ ei
– Ymi = Yi + ei = a Xi + ei
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 18 of 35
Example data
Measured Deflection under load
60
50
mm
40
30
20
10
0
0
2
4
6
8
10
12
Kilos
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 19 of 35
Regression Analysis – Math

Object: find model that gives “best fit” to data
– That is, model that minimizes total “error”




Total Error ≡ Sum of squared errors = ∑ (ei )2
Why does this make sense?
Because this gives a Gaussian distribution,
that is, bell-shaped or random
Solution Concept:
– Optimization criteria: δ(error) / δ(parameter) = 0
– Solve for each parameter of model (ex: “a”)
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 20 of 35
Regression Analysis – Graph
Model of Deflection under Load
60
50
mm
40
30
20
10
0
0
2
4
6
8
10
12
Kilos
From Excel using LINEST Formula
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 21 of 35
Regression Analysis - Practice
A pause for Demonstration
of Excel analysis you can use
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 22 of 35
Note: Lab work versus Real World

Object: “best fit” of data to function Y = f(X)

If, as in lab, you control all factors and vary
a few, then you can be clear on independent
and dependent variables and show causality
Example: weights on spring cause deflection
Thus you can write equation:
Deflection = a(Weight) + measurement errors

Is this true in “real world”, outside lab



Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 23 of 35
Real World





It is known that there is a close correlation, a
good statistical fit, between number of firemen
at a fire and amount of fire damage:
Fire damage = f (number of fireman)
What does it say about effect of firemen?
Nothing much! In this case, correlation is
“spurious” :
Size of fire => Damage and Number of Firemen
thus both also related
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 24 of 35
Consider Effect of Different Periods

The Data are all for the price of Google stock
(GOOG) starting in August 2004

First, the results for 3 years ending Sept 2007
Next, 3 years ending Sept 2008 DIFFERENT!
How about 4 years 2004-08 YET ANOTHER!

Results depend on set of data !!!


Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 25 of 35
Best Fit GOOG 04 - 07
Best Fit GOOG 04 - 07
700.00
600.00
Price ($)
500.00
400.00
300.00
200.00
100.00
Original Data
Best Fit
r = 3.54 % /month
0.00
0
5
10
15
20
25
30
35
40
Time (months starting 08/2004)
Original Data
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Best Fit
Richard de Neufville ©
Uncertainty assessment
Slide 26 of 35
Best Fit GOOG 05 - 08
Best Fit GOOG Prices 05-08
800.00
700.00
Price ($)
600.00
500.00
400.00
Best Fit
r = 1.43 % month
300.00
200.00
Original Data
100.00
0.00
13
18
23
28
33
38
43
48
53
Time (months starting 09/2005)
Original Data
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Best Fit
Richard de Neufville ©
Uncertainty assessment
Slide 27 of 35
Best Fits GOOG
Best Fits GOOG
800.00
Best Fit 04 – 07
r = 3.54% / month
700.00
Price ($)
600.00
500.00
400.00
300.00
200.00
100.00
Original Data
Best Fit 04 -08
R = 2.79% / month
0.00
0
10
20
30
40
50
Time (months starting 08/2004)
Original Data
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Months 0-48
Months 0-36
Richard de Neufville ©
Uncertainty assessment
Slide 28 of 35
Revision of Estimates –
Bayes Theorem

Definitions
– P(E)
Prior Probability of Event E
– P(E/O) Posterior P(E), after observation O is
made. This is the goal of the analysis.
– P(O/E) Conditional probability that O is
associated with E
– P(O)
Probability of Event (Observation) O


Theorem: P(E/O) = P(E) {P(O/E) / P(O) }
Note: Importance of revision depends on:
– rarity of observation O
– extremes of P(O/E)
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 29 of 35
Application of Bayes Theorem

At a certain educational establishment:
P(students) = 2/3
P(fem/students) = 1/4

P(staff) = 1/3
P(fem/staff) = 1/2
What is the probability that a woman on
campus is a student?
{i.e., what is P(student/fem)?}
P(fem/student)
P(student/fem) = P(student)
P(fem)

Thus: P(student/fem) = 2/3 {(1/4) / 1/3)} = 1/2
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 30 of 35
Likelihood Ratios, LR
Definitions
P(E )
LR
P(E)
= P(E does not occur)
= > P(E) + P(E ) = 1.0
_
= P(E)/P(E ); therefore
= LR / (1 + LR)
LRi
= LR after i observations
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 31 of 35
Likelihood Ratios (2)

Formulas
LR1
= P(E) {P(Oj/E) / P(Oj)}
P(E ) {P(Oj/E ) / P(Oj)}
after a single observation Oj
CLRi
LRN = LR0
= P(Oj/E) / P(Oj/E )
the conditional likelihood ratio for Oj

Nj
(CLR
)
j
j
Nj = number of observations of type Oj
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 32 of 35
Application of LR (1)


Bottle-making machines can be either OK
or defective
P(D) = 0.1
The frequency of cracked bottles depends
upon the state of the machine
P(C/D) = 0.2
P(C/OK) = 0.05
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 33 of 35
Application of LR (2)
Picking up 5 bottles at random from a machine,
we find {2 cracked, 3 uncracked}
What is the Prob (machine defective)
LRO = P(D) / P(OK) = 0.1/0.9 = 1/9
CLRC = 0.2/0.05 = 4
CLRuc = 0.8/0.95 = 16/19
LR5 = (1/9) (4)2 (16/19)3 = 1.06
P(D/{2C, 3UC}) = 0.52 = 1.06/(1 + 1.06)
Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 34 of 35
Take-aways from today



Many ways to try to estimate uncertainties
None are clear winners, each has its own use
Statistical analysis may be best, a good way
to make sense out of available data

HOWEVER, each method has its difficulties
Statistical analysis concept “house of cards”
Above all: Correlation does not prove Cause

Bayes Theorem – More on this later


Engineering Systems Analysis for Design
Massachusetts Institute of Technology
Richard de Neufville ©
Uncertainty assessment
Slide 35 of 35