You are on page 1of 7

# Preference Logic

Pramod Parajuli
2011
Preference Logic Pramod Parajuli, 2011
1
Preference Logic Pramod Parajuli, 2011
2
0
~C
1
~C
2
~~C
997
~C
998
~C
999

Solution: use of Just Noticeable Dierence ( JND)

C
i
`C
j
i, u(C
i
)-u(C
j
) `
> 0
C
0
~C
1
~C
2
C
997
~C
998
~C
999

Preferences are states of mind whereas choices are actions.
Pramod Parajuli, 2011 Preference Logic
3
The problem domain
! Optimal solutions are dicult to obtain.
! Finding sub-optimal solutions
Relaxing the objective function or
Relaxing the constraints
Pramod Parajuli, 2011 Preference Logic
4
Agenda
! To explore preference logic
Pramod Parajuli, 2011 Preference Logic
5
Preference types
! Label preference, e.g. nite set and alternatives
! Object preferences, e.g. orange vs. apple
! Action-object preference, e.g. drinking tea or drinking coee
! and many other types: intrinsic, extrinsic, conditional
preferences etc.
Pramod Parajuli, 2011 Preference Logic
6
Preference operators
! Preference operator:
e.g. A`B
transitive, if A`B and B`C, then A`C
! Indierence: ~
e.g. A~B
reexive, if A~B, then B~A
! At least as good as:
e.g. AB
transitive, reexive
Pramod Parajuli, 2011 Preference Logic
7
Basic preference operator
! At least as good as:
AB
! Preference:
A`B i AB and (BA)
! Indierence: ~
A~B i AB and BA
Pramod Parajuli, 2011 Preference Logic
8
Choice
! Choice is revealed preference (action).
! Choice function
Let C be a choice function.
If C is applied for set of alternatives B then
for all BCA: C(B)CB,
if B, then C(B)
Pramod Parajuli, 2011 Preference Logic
9
A
B
Choice properties (social)
! property (Cherno)
If BCA then B1C(A) CC(B)

Problem: C({A,B,C}) = {A}
C({A,B}) = {B}
B =
! property
If BCA and X,YC(B), then XC(A) i YC(A)

Problem: One must be Australia champion to become world champion.
! ! - property (expansion)
C(A
1
)11C(A
n
) CC(A
1
A
n
)
Pramod Parajuli, 2011 Preference Logic
10
Choice properties (economic)
! Weak axiom of revealed preference (WARP)
If X,YA and XC(A), then for all B,
if XB, and YC(B), then XC(B)

Problem: oensive choices
! Strong axiom of revealed preference (SARP)
Recursive closure of WARP
In words: From a set of alternatives A
1
, if X is chosen while Y is available, and
if in some other sets alternatives A
2
, Y is chosen while Z is available, then there
can be no set of alternatives containing alternatives X and Z for which Z is
chosen but X is not.
Pramod Parajuli, 2011 Preference Logic
11
! Good: better than its negation
A ` A
! Bad: worse than its negation
A ` A
! Goodness
B. A ` B~B
B. B~B ` A
Pramod Parajuli, 2011 Preference Logic
12
Preference metrics
! Completeness (incompleteness)
! Transitivity
! Order
A
i
`X
k
`A
j
or A
j
`X
k
`A
i
holds for each pair of labels (A
i
, A
j
), ij
Pramod Parajuli, 2011 Preference Logic
13
Constructing preferences from choice
ree dierent methods:
i. At least as good as
XY i for some B, XC(B) and YB
X`Y i XY and (YX)
XY i XY and YX
ii. At least as good as in a binary set
XY i XC({X,Y})
X`Y i XY and (YX)
XY i XY and YX
Pramod Parajuli, 2011 Preference Logic
14
Constructing preferences from choice
iii. Strictly preferred to
X`Y i for some B, XC(B) and Y[B\C(B)]
XY i (X`Y)
XY i XY and YX

Pramod Parajuli, 2011 Preference Logic
15
Learning preferences
! Lets consider an agent is able to choose among worlds
! Lets consider two propositions dene possible set of worlds:
p e agent mostly visits Bondi beach.
q e agent plays skate on the way.
Now, four possible set of worlds can be dened:

W
1
: p.q
W
2
: p.q
W
3
: p.q
W
4
: p. q
Pramod Parajuli, 2011 Preference Logic
16
Learning preferences
! Now, lets consider, through interaction with many agents, we
found probability and desirability of the possible set of worlds
World Probability Desirability
W
1
: p.q 1/6 -2
W
2
: p.q 2/6 1
W
3
: p.q 2/6 -1
W
4
: p. q 1/6 3

e value of a proposition can be evaluated as:
value = for all true occurrence/s of proposition, sum(probability desirability)
Pramod Parajuli, 2011 Preference Logic
17
Learning preferences
! value of p, #(p) =
! value of q, #(q) =
! Similarly, #(p) = , #(q) =
! Since #(p)>#(q) and #(q)>#(p), we conclude that pq.
e agent prefers going Bondi beach than playing skate.
Pramod Parajuli, 2011 Preference Logic
18
1
6
2
#
\$
%
&
'
(+
2
6
1
#
\$
%
&
'
( = 0
1
6
2
#
\$
%
&
'
(+
2
6
1
#
\$
%
&
'
( =
4
6
1
6
5
6
Implementation
! Implementation of preferences has been taken as relaxation in
optimization procedure ( Jayaraman & Govindrajan & Mantha,
1998).
! It is achieved through constraint relaxation.
Lets consider, we have a function shortest-path dened as;
shortest-path(X, Y, C, P) path(X, Y, C, P). path P with distance C from
X to Y.
shortest-path(X,Y,C
1
,P
1
) shortest-path(X,Y,C
2
,P
2
) C
1
< C
2
.
Pramod Parajuli, 2011 Preference Logic
19
Challenges
! Expressivity and completeness
! Transitivity
! Dierent world problem, preference change, temporal preferences
! Belief
! Commitments
! Privacy
! Criticisms (People do and should act as problem solvers, not maximizers, because
they have many dierent and incommensurable goals to achieve Steven G. Kranz)
Pramod Parajuli, 2011 Preference Logic
20
ank you for your kind attention.

uestions and suggestions are welcome.
Preference Logic Pramod Parajuli, 2011
22
Bibliography
e primary source of the concepts presented here is the article Preferences in
Stanford Encyclopedia of Philosophy, http://plato.stanford.edu/entries/preferences/
Accessed on August 12, 2011.

Preference Logic Pramod Parajuli, 2011
23
Annex-1: constructing choice from preferences
ree cases:
i. e best choice connection
C
B
(B) = {X B | `Y B: (X`Y)}
ii. e non-dominance choice connection
C
L
(B) = {X B | `Y B: (Y`X)}
iii. e optimization choice connection
When cyclic preferences exist, A`B`C`A, C
B
(A,B,C) = C
L
(A,B,C) =
Pramod Parajuli, 2011 Preference Logic
24
Annex-1: constructing choice from preferences
iii. e optimization choice connection continued
Let S be the set of these sets. Now, B is in S i:
a) B r(A)
b) for all X,Y: if XA\B and YB then (X `Y)
c) for all FB there is a YF such that for some XA\F: X`Y
Now, the choice function is dened as the union of S:
C
O
(A) = S
Pramod Parajuli, 2011 Preference Logic
25
Annex-2: details of SARP
If
X
1
,X
2
, ..., X
n
A
1
,
X
2
, ..., X
n
A
2
, ...,
X
n-1
,X
n
A
n1
,
X
n
A
n
, and
X
1
C(A
1
), X
2
C(A
2
), ..., X
n
C(A
n
),
then,
for all B with X
1
,X
2
,...,X
n
B, if X
i
C(B), i{1,...,n},
then X
1
,X
2
,...,X
i1
C(B) (SARP)
Pramod Parajuli, 2011 Preference Logic
26
Annex-3: questions raised at the end
Any formal model for desirability?
How would we capture desirability? ualitative desirability?
How preferences would t in automated planning? How would the cost model
developed?
Can preferences contribute to risk-modeling and mitigation?
How can the machine-learning methods be used for capturing preferences?
How preference-based model diers from Bayesian decision making theory?
Pramod Parajuli, 2011 Preference Logic
27