axioms
DESCRIPTION
aaaaaTRANSCRIPT
-
PMR 2728 / 5228Probability Theory in AI and
Robotics
Subjectivist AxiomsFabio G. Cozman - Ofce MS08
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.1/36
-
Foundations: decision making
First step is to define acts.1. An act is a function X :
-
Utility theory
1. Preferences amongst consequences2. In fact, preferences amongst lotteries3. Basic lottery: (r1, , r2) yields combination of r1 and r2
with weight 4. Obvious interpretation:
(r1, , r2) = r1 + (1 )r2.
5. That is, lotteries are convex combinations
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.3/36
-
von Neumann-Morgenstern theorem
Theorem 1. SupposeR is a set of lotteries, and is a binary relationamong lotteries such that:
1. the relation is complete and transitive;2. if l1 l2 and (0, 1), then (l1, , l3) (l2, , l3);3. if l1 l2 l3, then there exists , (0, 1) such that
(l1, , l3) l2 (l1, , l3);then there is a function u : R < such that
l1 l2 u(l1) u(l2),
u((l1, , l2)) = u(l1) + (1 )u(l2)
and u is unique up to a linear transformation au + b with a > 0.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.4/36
-
What does it show??1. A rational agent has a utility function u2. Utility is unique up to linearity3. Decision making without uncertainty:
X1 10X2 0X3 5
X1 dominates the others, should be selected
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.5/36
-
Preferences amongst acts
Consider several acts:
1 . . . j . . . n
X1 c11 c1j c1n...
Xi ci1 cij cin...
Xm c1m cmj cmn
How to compare them?
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.6/36
-
Example
Take a ball from an urn.Urn with 10 red balls, 20 yellow balls, 30 black balls.
Red Yellow BlackX1 100 0 0X2 100 100 100X3 0 0 100X4 0 100 100
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.7/36
-
Ellsbergs example
Take a ball from an urn.Urn with 30 red balls and 60 yellow/black balls.
Red Black YellowX1 100 0 0X2 0 100 0X3
X4
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.8/36
-
Ellsbergs example
Take a ball from an urn.Urn with 30 red balls and 60 yellow/black balls.
Red Black YellowX1 100 0 0X2 0 100 0X3 100 0 100X4 0 100 100
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.9/36
-
Rules of the game
1. There is a relation amongst acts2. Axioms on :
Partial Order is reflexive and transitive.Independence/Sure-thing X Y iff Z, (0, 1),
X + (1 )Z Y + (1 )Z.
Dominance If X() > Y () for all , then
X Y and not Y X.
Continuity If Xi X and Y Xi Z then Y X Z.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.10/36
-
Transitivity
Take a ball from an urn.Urn with 10 red balls, 20 yellow balls, 30 black balls.
Red Yellow BlackX1 100 0 0X2 0 100 0X3 0 0 100
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.11/36
-
Independence/Sure-thing
Take a ball from an urn.Urn with 10 red balls, 20 yellow balls, 30 black balls.
Red Yellow BlackX1 100 0 0X2 0 100 0X3 100 0 100X4 0 100 100
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.12/36
-
Dominance
Take a ball from an urn.Urn with 10 red balls, 20 yellow balls, 30 black balls.
Red Yellow BlackX1 0 0 0X2 100 100 100X3 0 0 0X4 0 0 100
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.13/36
-
Back to Ellsbergs example
Take a ball from an urn.Urn with 30 red balls and 60 yellow/black balls.
Red Black YellowX1 100 0 0X2 0 100 0X3 100 0 100X4 0 100 100
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.14/36
-
Direct consequences of axioms:
1. X Y iff X Y 0.2. X 0 and Y 0 imply X + Y 0.3. X 0 then X 0 for > 0.4. X Y then X Y for > 0.
That is, there is a cone of desirable acts.Note that if X 0, then 0 X.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.15/36
-
Cones1. There is a cone of desirable acts.2. There is a cone of undesirable acts.3. There is a region of indecision (if X is there, X is
there).
How do we represent such cones?
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.16/36
-
Main theoremTheorem 2 (Giron and Rios 1980). If satisfies the axioms, then thereis a credal set K such that:
X Y P K : EP [X] EP [Y ]
Moreover, the maximal credal set is convex.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.17/36
-
Example
1. Binary possibility space2. X = [1, 2]3. Y = [3, 0]4. X Y5. Then E[X] E[Y ] implies
1p1 + 2p2 3p1
p2 p1,
and also p1 + p2 = 1.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.18/36
-
Continuing...
1. X = [1, 2], Y = [3, 0]2. X Y ; then p2 p1.3. 5Y/3 X4. Then 5E[Y ] /3 E[X] implies
5p1 1p1 + 2p2
p2 2p1,
and also p1 + p2 = 1.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.19/36
-
Geometry
1. A general assessment on the expectation for a variablerepresents a hyperplane on desirable acts.
2. Each expectation produces a constraint on probabilitymeasures.
3. A convex set of probability measures represents a setof desirable acts.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.20/36
-
Completeness
1. Additional axiom: is complete (either X Y or Y Xor both.
2. Then:(a) Suppose we have a credal set with P1 6= P2
representing .(b) Suppose EP1 [X] > EP2 [X].(c) Take = (EP1 [X] + EP2 [X])/2.(d) Then EP1 [X ] = (EP1[X] EP2[X])/2 > 0(e) Then
EP1[X ] > 0 > EP2 [X ]
and this contradicts completeness (neither X nor X).
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.21/36
-
Bayesian theory: Basic theorem
Theorem 3. Preferences satisfying partial order, independence,dominance, continuity and completenessare represented by a single probability measure.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.22/36
-
Other axioms for Bayesian theory
1. Savages theory: obtain utility and probability fromaxioms.
2. Anscombe/Aumanns: obtain utility and probability,assuming chances but single ordering .
3. Dutch book arguments (Ramsey, de Finetti)
(Also, some more general theories; SSK95, etc.)
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.23/36
-
Dutch book: basics
1. Theorem of the alternative: basic result in linearalgebra, with many versions.
2. Consequence of the theorem of the alternative:Theorem 4. A is a matrix, p a vector, a vector. Then exactly oneof the following systems has a solution:
Ap = 0, p 0,
i
pi = 1.
A < 0.
3. Suppose each row of A is (Xi i).4. Now we interpret: there is a probability measure p such
that E[Xi i] = 0 iff A < 0 has no solution.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.24/36
-
Dutch book
1. So, when the i fail expectation axioms, there is no p,then
[1 . . . n]
X1 1...
Xn n
< 0
has a solution.2. It means that it is possible to form an act that will be
always against you...3. If one can create such an act, one can make a Dutch
book you always lose the game!
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.25/36
-
Dutch book: ConclusionEither:you have a probability measure representing yourpreferencesorsomeone will create a Dutch book against you (and you willbe miserable)
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.26/36
-
Conditional preferences
1. For event B define B as
X B Y XB +Z(1B) Y B +Z(1B), any Z.
2. Given independence axiom, Z can be suppressed:XB Y B is enough.
3. Thus, if B(X Y ) 0 then X Y is desirable given B.4. The relation B satisfies all axioms ONLY IF P(B) > 0!!
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.27/36
-
Conditional expectations
1. If P(B) > 0, then
X B Y B(X Y ) 0 E[B(X Y )] 0.
2. Then
E[XB] E[Y B] E[X|B] P(B) E[Y |B] P(B)
and then
X B Y E[X|B] E[Y |B] .
That is, conditional expectations represent conditional pref-erences!
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.28/36
-
A few advanced issues
Conditioning can be defined even on events ofprobability zero.Binary preferences lead to convexity; more generalcredal sets can be produced with general choice.
We will not look into these issues any further in this course.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.29/36
-
Problem with zero probabilities
1. Suppose conditioning event B with probability zero.2. Then E[XB] E[Y B] for any X, Y !3. Now suppose (XB)() > (Y B)() for every B; this
violates dominance for B!!
There are theories that allow conditioning even if P(B) = 0.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.30/36
-
Universal conditioning
1. Instead of dening
P(A|B) =P(A B)
P(B),
we adoptP(A B) = P(A|B)P(B)
as an additional axiom.2. Now B can have probability zero.3. de Finettis theory, also known as Popper measure
theory; popular in some specialized fields, but notwidely used.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.31/36
-
Convexity?
1. So far, only binary choices.2. Maximal credal sets are convex.3. But convexity is not really required.4. Are there decisions that move away from convexity?
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.32/36
-
Sets of acts1. Consider 4 acts in a binary possibility space:
W = [2/5, 2/5],X = [0, 1],Y = [1, 0],Z = [0, 0].
2. Assume P(1) [1/4, 3/4].3. Note that Z is never the best.4. So, Z is dominated and discarded.5. How about the others?6. All are undominated (named maximal).7. Only X and Y can be optimal (E-admissible).
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.33/36
-
Walleys theorem
Theorem 5. Suppose set of acts is convex. Then all maximal acts areE-admissible.
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.34/36
-
Seidenfelds axiom1. Now suppose we require that decisions should not
change if the set of acts is convexified.2. Then, every decision must be an E-admissible
decisions!
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.35/36
-
General credal sets1. With Seidenfelds axiom, we can use sets of acts to
differentiate between any two credal sets!2. That is, any credal set represents a unique pattern of
preferences...3. This is the most general theory that has a behavioral
interpretation (that is, a theory that really makes anysense).
PMR 2728 / 5228Probability Theory in AI and Robotics Subjectivist Axioms p.36/36
Foundations: decision makingUtility theoryvon Neumann-Morgenstern theoremWhat does it show??Preferences amongst actsExampleEllsberg's exampleEllsberg's exampleRules of the gameTransitivityIndependence/Sure-thingDominanceBack to Ellsberg's exampleDirect consequences of axioms:ConesMain theoremExampleContinuing...GeometryCompletenessBayesian theory: Basic theoremOther axioms for Bayesian theoryDutch book: basicsDutch bookDutch book: ConclusionConditional preferencesConditional expectationsA few advanced issuesProblem with zero probabilitiesUniversal conditioningConvexity?Sets of actsWalley's theoremSeidenfeld's axiomGeneral credal sets