Vous êtes sur la page 1sur 55

Belief Revision in Non-Classical Logics

Pankaj Kumar Singh(Y0222)


Department of Mathematics
Indian Institute of Technology
Kanpur, INDIA
M.Sc. thesis
12 April, 2005
Contents
1 Introduction 5
1.1 Epistemic States and Belief Sets . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Logical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Expansion, Contraction and Revision 8
2.1 Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Revision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Motivation Behind Postulates . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Interrelations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Constructing Contraction and Revision 14
3.1 Epistemic Entrenchment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Partial Meet Contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 System of Spheres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4 Implementation of Belief Revision . . . . . . . . . . . . . . . . . . . . . . . 20
3.4.1 Finite Partial Entrenchment Ranking . . . . . . . . . . . . . . . . . 21
3.4.2 A PROLOG Implementation . . . . . . . . . . . . . . . . . . . . . . 22
4 Three-Valued Logic and Belief Revision 24
4.1 The Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 A Sequent Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2.1 Truth of a Sequent . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.2 Provability of a Sequent . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.3 Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Automated Proof Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.1 Semantic Tableaux . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.2 Tertiary Decision Diagrams . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 Contraction and Revision . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.5 Interrelations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.6 Epistemic Entrenchment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.7 Partial Meet Contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.8 System of Spheres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.9 Interpretation and Applicability . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.10 Generalization to n-valued Logics . . . . . . . . . . . . . . . . . . . . . . . . 35
2
CONTENTS 3
5 Rough Belief Revision 37
5.1 Rough Truth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.1.1 The semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2 Rough belief change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3 Partial Meet Contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.4 Rough Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.5 Maximal Consistent Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.6 Revision from System of Spheres . . . . . . . . . . . . . . . . . . . . . . . . 45
5.7 Epistemic Entrenchment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.8 Interrelations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6 Conclusions 52
6.1 Scope for Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Acknowledgements
I would like to take this opportunity to thank my guide Dr. Mohua Banerjee for her support
and guidance throughout this project and her patience at some trying times. I would also
like to thank her for providing the initial motivation, which led me to take up this topic
as a reading project and later as my M.Sc. project. I also thank all my friends and wing
mates for being such good company that I could always refresh my eorts at tackling the
challenges posed by this project.
Pankaj Kumar Singh
4
Chapter 1
Introduction
Belief revision models the change in beliefs of a person when some new information is
provided or an old belief has to be retracted. This kind of study helps in building intelligent
agents like robots, who have to manage beliefs about the world in order to achieve their
goals. We all know that beliefs can sometimes be wrong, so intelligent agents should be able
to revise beliefs when new correct information is acquired. When giving up a belief one has
to decide which of its consequences to retain. Thus a fundamental issue in belief revision is
how to decide, which information has to be retracted in order to avoid contradiction. Other
issue is that logical consequences alone do not tell which beliefs to give up. What do we
choose, when contradiction can be avoided by retracting one of the two given beliefs?
The problem of belief revision is very well studied in classical background, where any
statement is either true or false. There are well-dened functions which model the belief
change and give out the changes explicitly. But gradually it is realized that belief change is
not that naive. In real world settings an agent is bombarded with all sorts of information,
that cannot be rejected or accepted decisively. So there have to be some beliefs which are
weakly held or they state vague truth. In any case we have to incorporate vagueness, which
can be achieved by weakening the background classical logic.
In this thesis we work on two methods of changing the background logic. First is that
we insert more truth values into classical logic. More truth values increase the variety
of beliefs and hence increase vagueness notions. Other way is to decrease the credibility
of existing beliefs by including incomplete information as beliefs. With these motivations
we investigate 3-valued logic in which there is an extra truth value and rough logic where
incomplete information is incorporated into the formal system.
After investigating the 3-valued logic we realized that change is only with respect to
interpretation of new beliefs, which more or less satisfy the same rules or postulates of belief
change. We expect similar results with general n-valued logics. But for the rough logic we
realize that rules of belief change do not t well and require modications. Currently we
are working on explicit denition of belief change functions as in the classical case and also
thinking of right postulates of belief change with respect to rough logic. In the remaining
chapter we present a formal treatment of preliminaries for belief change.
1.1 Epistemic States and Belief Sets
Epistemic states are the central entities in the belief change theory as they represent an
actual or a possible cognitive state of some individual at a given point of time. Our notion
of epistemic states is idealization of psychological concepts like consistency of the state
5
6 CHAPTER 1. INTRODUCTION
and closeness under logical sequences. Since a persons beliefs tend to be in equilibrium
state as regulated by above two conditions, its interesting to see how they alter themselves
when disturbed by contradictory notions. Before entering into dynamics of these changes
we introduce propositional model for epistemic state called as belief sets. In this model
epistemic states are sets of propositions, expressed by sentences from some given object
language. The intended interpretation of such a model is that the set consists of those
propositions that the person accepts in the present state. The central criterion of change is
consistency and logical closure.
1.2 Logical Preliminaries
We already know that belief set is a propositional model of an epistemic states where a
proposition or sentence in a belief set means that the individual believes it to be true or
regards it as certain. This model is linguistic and presupposes a language L. One can
assume that L consists of expressions formed by sentential connectives : Negation : ,
Conjunction : , Disjunction : , Implication :. The symbols and stand for the
sentential constants falsity and truth. Formally our epistemic states are based on set
of all propositional well formed formulae (w). Propositional ws can be described by
following inductive denition:
(i) Any propositional letter p which can be directly assigned truth values T, F is a w.
(ii) If A and B are ws then A, A B, A B and A B are ws.
We use L to denote set of all ws. For dening concepts in belief change we use classical
logic and present the equivalent notions whenever talking in other logics. After realizing
our language we have to dene logical consequence. In classical background we dene an
entailment relation between subsets of L and L. This relation is syntactic inference rela-
tion. Similar to there is a binary relation [= based on semantics, if [= then whenever
all propositions in are T, is T. We know the (completeness and soundness):
i [=
Other deduction properties like:
compactness-
If , where L, then
1
...
n
s.t.
1
, ...,
n

weakening-
If and then
deduction theorem-
If then
modus ponens-
If and then
1.2. LOGICAL PRELIMINARIES 7
, are known to us. With language L and inference relation we can dene logical closure
Cn() of a subset as
Cn() = L :
We dene theories of L as subsets satisfying
Cn() =
From here we can assume that belief sets (generally denoted by K) are theories of L. The
largest belief set, denoted as K

is the set of all beliefs or propositions. The falsity


constant entails everything and hence Cn() = K

. Similarly truth constant is


entailed by every proposition, meaning is in every belief set. In addition any belief set
contains all theorems (ws s.t. or ). We call subsets of L consistent if -
B L s.t. and
and inconsistent otherwise. Note that by the denition any set containing is
inconsistent. Following result is very useful in classical logic-
If is inconsistent then
Notion of maximal consistent sets is also useful, a subset M of L is maximal consistent
if it is consistent and / M implies M is inconsistent. Every consistent set has a
maximal consistent extension.
Chapter 2
Expansion, Contraction and
Revision
2.1 Expansion
Expansion models epistemic change when something new is learned [8],[6]. The incoming
information may produce new connotations leading to inconsistencies, making the new epis-
temic state inconsistent. The most common causes of such changes are observations and
the information provided by other people. Thus expansion + takes some belief set K /
L
(/
L
is set of all belief sets) and the new information A L to another belief set K + A.
The expected nature of a expansion function is regulated as following postulates-
(K
+
1) K +A is a belief set
(K
+
2) A K +A
(K
+
3) K K +A
(K
+
4) If A K then K = K +A
(K
+
5) If K H then K +A H +A
(K
+
6) For a belief set K and sentence A, K+A is the smallest belief set that satises
(K
+
1) (K
+
5)
One can easily show that (K
+
1) (K
+
5) imply Cn(K A) K + A and (K
+
6) is
equivalent to K + A Cn(K A). Thus all the six postulates enforce a explicit deni-
tion of expansion function as -
K +A = Cn(K A) (2.1)
2.2 Revision
This type of belief change comes into picture when the learned information A contradicts
the beliefs in current epistemic state K. To maintain consistency one needs to revise his/her
beliefs by removing any thing incompatible with A. Thus attempt to avoid inconsistency
during addition of new information distinguishes a revision function from expansion. For-
mally, a revision function * takes a belief set K /
L
and the new information A L to
another belief set K A satisfying following postulates-
8
2.3. CONTRACTION 9
(K

1) K A is a belief set
(K

2) A K A
(K

3) K A K +A
(K

4) If A / K then K +A K A
(K

5) K A = K

if and only if A
(K

6) If A B then K A = K B
(K

7) K (A B) (K A) +B
(K

8) If B / K A then (K A) +B K (A B)
Above postulates dont explicitly tell which beliefs to be retained and which to be removed.
They only circumscribe the set of rational solutions and dont lead to a unique solution as
in case of expansion.
2.3 Contraction
The change in epistemic state is called as contraction when some belief A has to be retracted
without adding extra information. Contraction as the name suggests requires removal of
all those beliefs which can imply the currently held belief A. A contraction function must
satisfy following postulates-
(K

1) K A is a belief set
(K

2) K A K
(K

3) If A / K then K A = K
(K

4) If ,A then A / K A
(K

5) K (K A) +A
(K

6) If A B then K A = K B
(K

7) (K A) (K B) K (A B)
(K

8) If A / K (A B) then K (A B) K A
Similar to revision, the postulates are not sucient to produce a unique contraction func-
tion and a notion of preference among the beliefs is necessary. Like for retracting A B
from CnA B, removing either A or B is sucient. Thus a sense of preference is must
to make a choice among the two possibilities.
2.4 Motivation Behind Postulates
Postulates for revision and contraction, as already said only regulate the desired properties
of these functions and dont lead to explicit denitions. K

1 6 are elementary require-


ments that connect K, A and K A and are called basic set of postulates. K

7 and K

8
are concerned with minimal change. To ensure minimal change in K when revised with
A, expansion(+) of K A with another formula B and revision of K directly with A B
must be same provided K A + B does not become K

. Following alternative postulates


for revision look more closer to K

7 and K

8:
K

K A K B K A B
K

If A / K A B then K A B K A
Lemma 2.4.1 (i) K

7 is equivalent to K

and (ii) K

8 is equivalent to K

.
10 CHAPTER 2. EXPANSION, CONTRACTION AND REVISION
Proof (i) Suppose K

7 is true. Let K A K B, using K A = K (A B) A


K AB+A we get A K AB. Similarly B K AB. But that is same
as A B K A B. Since A B K A B, K A B.
Let K

holds. If KAB then B KAB. Since B KAB and


B (B ), B KAB. K

now gives B K(AB)(AB) =


K A. Hence K A+B.
(ii) Suppose K

8 holds. Taking A

= A B and B

= A in K

8 directly gives K A
B +A K A, where K A B K A B +A.
Suppose K

holds. Let B / K A = K (AB) A, hence A B (AB) /


K (AB) A. Taking A

= AB and B

= A in K

we get K A K AB. Thus


K A+B K A B +B = K A B.
Following also follow from K 7, 8 and are useful sometimes:
(i) K A B = K A or K A B = K B or K A B = K A K B
(ii) K A = K B iff B K A and A K B
Proof (ii) only if follows directly. So for if part, suppose B K A and A K B. If A
or B is a theorem then K A = K B = K

. So suppose , A and , B. Thus K A


and K B can not be K

and hence B / K A and A / K B. But then using K

8,
K A = K A+B = K A B = K B +A = K B.
Similar to revision postulates K

16 are basic. If A K then K

14 give KA+A
K. Thus with K

5 (recovery), (K A) +A = K whenever A K.
Lemma 2.4.2 K

7 is equivalent to
K

K A Cn(A) K A B
Proof Let K

7 holds. Suppose KACn(A). If A / K then KA = KAB = K


and hence there is nothing to prove. So we assume A K and show that K(A B),
so that K A K (A B) K A (A B) = K A B. Using recovery we
get (A B) A K (A B), but then A (A B) A implies A K (A B)
and Cn(A) gives K (A B).
Let K

holds. Let K A K B. Since K A and A CnA, we get


A K ACnA K AB. Similarly B K AB. Thus (AB)
(A B) K A B. Using recovery we also have (A B) K A B, and
hence K A B.
Now we prove a result regarding equality of contraction functions.
Theorem 2.4.1 If B A K A and A B K B then K A = K B. If
A, B K then K A = K B implies B A K A and A B K B.
Proof We assume B A KA and A B KB and prove that KA = KB =
KAB. If A or B KAB then AB KAB which implies KA = KB = K
( A B). Thus we consider the case when A, B / K A B, hence by K

8 we have
KAB KA, KB. So we only need to prove KA, KB KAB. Let KA.
Using recovery and assumption we have (A B) K A and (A B) K B.
Thus using K

7 we get (A B) K (A B). Using recovery again on K A B


with K we have (A B) K A B. Since A and B are equivalent in
K A B we have A K A B. Now we prove that A A is also in
2.5. INTERRELATIONS 11
K A B, so that we get K A B. Now if A / K then A B / K which implies
K = KAB KA, KB, that is to say KA = KB = K. So we take A K. Us-
ing recovery we get (A B) A A K(A B). Thus A K(A B), KA.
K

7 now gives A K (A B) A = K A B. Hence K A K (A B).


Similarly K B K (A B) and we have proved K A = K B = K A B.
For the second part, taking A, B K and recovery gives that A B KA = KB.
Using the above result we prove following factoring condition for contractions.
K A B = K A or K A B = K B or K A B = K A K B
Proof Taking A

= A B and B

= A in previous lemma, we have, if A (A B)


K A B and A B A K A then K A B = K A. But since A B A
and A (A B) A B, our condition modies to: if A B K A B then
K A B = K A. Now, if either A B K A B or B A K A B then
K A B = K A or K A B = K B. When both the conditions fail, that is
A B / K A B and B A / K A B then because B (A B) we have
B / K A B and A / K A B. K

8 then gives K A B = K A K B.
Observation 2.4.1 If A K then A K (A B). We have used this result several
times in above proofs and will use it again in next chapter.
Observation 2.4.2 If A K A B or B A K A B then K A B = K B.
A K A B gives B A K A B which means B (A B) K A B.
(A B) B being a theorem is in K B and hence using previous theorem we have
K A B = K B.
2.5 Interrelations
Changes in epistemic states modelled as expansion, revision and contraction are interrelated.
Revision of a belief set K with respect to a sentence A can be seen as removal of any
incompatible information A and then addition of A along with all logical consequences.
This relationship is called Levi-identity and is written as-
K A = (K A) +A (2.2)
Let us call the revision function above R(). Following theorem tells that function R()
dened through Levi - identity is indeed a revision function.
Theorem 2.5.1 If a contraction function satises (K

1) to (K

6) then R() satises


(K

1) to (K

6). Furthermore if also satises (K

7) and (K

8) then R() satises


(K

7) and (K

8) respectively.
Proof
(K

1) K A = (K A) + A is closed under logical consequences by 2.1 and hence is a


belief set.
(K

2) A K A as (K A) A A (dilution or weakening).
12 CHAPTER 2. EXPANSION, CONTRACTION AND REVISION
(K

3) From (K

2), K A K or (K A) A K A . Using weakening


Cn((K A) A) Cn(K A). 2.1 and 2.2 now give K A K +A.
(K

4) Let A / K, using (K

3) we have K A = K. Hence Cn(K A) = Cn((K


A) A) or K +A = (K A) +A.
(K

5) if Let A. Weakening and 2.1 give A and A (K A) + A. As B


L, A, A B, K A = (K A) +A = K

.
only if Let (K A) +A = K

.
Thus (K A) +A A and (K A) +A (A A).
Using D.T. K A (A A) and K A (A (A A)).
But second statement is same as K A ((A A) A) because (A
B) (B A). Using modus ponens we get K A A. (K

4) now implies
A.
(K

6) Let K A = (K A) +A. D.T. gives (K A) (A ). But A B


B A, using (K

6) gives (K B) = (K A) (A ). Weakening gives


(KB)+B (A ) which in turn yields (KB)+B as A (KB)+B.
Thus K A K B.
Similarly K B K A giving K A = K B.
(K

7) Let K A B = (K (A B)) + (A B).


(A B) (K (A B)) (D.T.)
Now A (A B) (A B)
By (K

7) and (K

6), K(AB)K(AB) K((AB)(AB)) =


K A.
If we show that (A B) K (A B) then proof will be complete as above
relation will give (AB) KA. But as ((AB) ) (A (B ))
applying modus ponens twice will yield ((K A) +A) +B = K A+B.
So (AB) K(AB) needs to be proved. Using (K

2) and (K

5) we have
K(AB) K K(AB) +(AB). Thus ((AB) ) K(A
B) +(A B) which on application of D.T. gives ((A B) ) K (A B) as
((A B) ) ((A B) ((A B) )).
(K

8) Let B / K A and K A + B = ((K A) + A) + B. Converse of D.T.


and D.T. along with A (A B) A give (A B) / K A and
(A B) K A = K (A B) A respectively.
Now using (K

8) with A

= A B and B

= A we get (AB) K(A


B) A K(A B). Finally using modus ponens K(A B) +(A
B) = K (A B) as (A B) (A B).
Similar to R(), contraction can be dened using a revision function. Contraction
represents knowledge left after removal of some particular information let us say A from
a belief set K. Equivalently, this can be seen as the information remaining inside K after
consistent addition of A to K. The consistent addition of A is essentially revision and
ensures removal of A from the knowledge base. This way of obtaining a contraction is called
Harper-Identity and is written as
K A = K K A (2.3)
We call the function above C() i.e. contraction C produced from a revision * .Fol-
lowing theorem ensures that C() is indeed a contraction function.
2.5. INTERRELATIONS 13
Theorem 2.5.2 If a revision function * satises (K

1) to (K

6) then C(*) satises


(K

1) to (K

6). Furthermore, if * also satises (K

7) and (K

8) then C(*) satises


(K

7) and (K

8) respectively.
Proof
(K

1) K A = K K A is a belief set as intersection of two belief sets is also a belief


set. If K K A then K, K A as K K A K, K A and both K
and K A are beliefs sets. Hence we have K K A.
(K

2) K A = K K A K.
(K

3) Let A / K or A / K. Using (K

4), K A = K + (A). But K K + (A)


giving K A = K K + (A) = K.
(K

4) Let , A and assume that A K A = K K A. Thus A K A. Using


(K

2), A K A and we have K A = K

. (K

5) now gives A or A
which is a contradiction and hence A / K A.
(K

5) Let K, we need to show (K K A) + A. (A ) implies


A K. Using Converse of D.T. its sucient to prove A K A. But
that is ne as A (A ) and A K A because of (K

2).
(K

6) If A B then A B. From (K

6) we have K A = K B which gives


K A = K K A = K K B = K B.
(K

7) K A K B which means K, K A, K B. Taking A

= (A B)
and B

= (A) and using (K

7) we get K ((A B) A) = K A K
(A B) + A. Similarly we can get K B K (A B) + B. Using
D.T., A , B K (A B). Now C() already satises (K

5) so
K (K K (A B)) + A B. And we have A B K (A B).
But A B , A , B and hence K (A B). Thus
K K (A B) = K (A B).
(K

8) Let A / K (A B). Which means A / K or A / K (A B). If A / K then


K(AB) K = KA. So consider the case when A K but A / K (AB).
Taking A

= A B, B

= A and using (K

8) we have K (A B) + A
K (A B) A = K A. But K (A B) K (A B) +A and we have
K (A B) = K K (A B) K K A = K A.
Chapter 3
Constructing Contraction and
Revision
In previous chapter contraction and revision function were investigated indirectly by means
of a number of postulates imposed on the functions. It was also shown in that if it is possible
to give explicit construction of a contraction function, then this also yields a revision function
by means of Levi-Identity (or vice-versa through the Harper-identity). Thus epistemic
change on belief sets is completely determined by contraction function (or revision function).
The goal of this chapter is to present three constructions for contraction function leading
to single proposal.
3.1 Epistemic Entrenchment
If one can rank beliefs such that a preference among the beliefs can be induced then a
contraction function can be constructed. One such ranking (an ordering) is called Epistemic
Entrenchment ranking. This ranking is particular to the belief set K under consideration
and is a binary relation satisfying following postulates-
(EE1) is transitive.
(EE2) If A B then B A
(EE3) Either A B A or A B B
(EE4) K ,= K

then B A for all B if and only if A / K


(EE5) If A B for all B then A
Now given an epistemic entrenchment ranking , contraction function can be dened as
K A =

B K : A < A B, if , A
K, otherwise
(3.1)
The symbol < is strict preference i.e. A < B (A B and A , B) A , B as is a
total order. A similar method can be used in the reverse way. Given a contraction function
an epistemic entrenchment ranking can be obtained as
A B if A / K (A B) or A B (3.2)
Following theorem proves that above ordering is indeed a epistemic entrenchment ranking.
14
3.1. EPISTEMIC ENTRENCHMENT 15
Theorem 3.1.1 Let be a contraction function and let K is a belief set. Then the ordering
dened in 3.2 is an epistemic entrenchment ranking.
Proof For the proof of EE(25)we will also consider an alternative of 3.2. We can replace
the premise A / K(AB) or AB by A / K(AB) or B K(AB). Using the
properties of contraction function, especially (K

4) one can check that the two denitions


are equivalent. Now we show that satises (EE1) (EE5).
(EE1) Suppose A B and B C. Thus (A / KAB or AB) and (B / KBC
or B C). So there are four possibilities:
(i) A / KAB and B / KBC. We prove that A / KAC. By contradiction
let us assume A KAC. By recovery ((AC) B) A K((AC) B).
But ((A C) B) A A and hence A K ((A C) B). Thus by K

7,
A K A C K ((A C) B) K (A B C). A K (A B C)
now gives B C / K (A B C) (if B C K (A B C) then by K

4
A B C which implies A K A B - a contradiction). From K

8 we now
get K (A B C) K B C. Since we assumed B / K B C, we get
B / K (A B C). Thus A B / K (A B C), and hence using K

8,
K (A B C) K A B. Now A K (A B C) K A B gives
contradiction. Thus A / K A C and A C.
(ii) A / KAB and BC. Suppose A KAC. Since BC, C KAC.
Hence A C K A C and we get A C. So A C.
(iii) AB and B / KBC. This is not possible as AB implies B KBC.
(iv) A B and B C. This directly implies A C and hence A C.
(EE2) We have to show that for all A, B L if A B then A B. Assuming A > B and
using the alternate denition we have A K (A B) and B / K (A B). But
then A B implies B K (A B) which leads to contradiction. Hence A B.
(EE3) We have to show for all A, B L, A AB or B AB. Let A , AB which
means that A KA(AB) = K(AB) and AB / K(AB). But then
B / K (A B) = K B (A B) must hold. Hence B A B.
(EE4) if Let A / K. Then (K

2) gives A / K (A B) and hence A B.


only if Since A B for all B L, A ( is such that B for all B L).
Which implies A / K (A ) = K () or K () as (A ) . Now
K K implies that K = K

and hence K is not possible. The


only possibility A / K , leads to A / K as K = K (K ,= K

/ K
K = K , using (K

3)).
(EE5) We have to show that if for all B L, B A then A. Letting B = we have
/ K (A ) = K A or A K A. First assertion cant be true and hence
second one must be true. Thus A K A, which using (K

4) gives A
Theorem 3.1.2 Let K be a belief set. For every contraction function of K there is an epis-
temic entrenchment (related to K) such that 3.1 is true. Conversely for every epistemic
entrenchment related to K, there is a contraction function satisfying 3.1.
Proof For the rst part theorem 3.1.1 gives the required epistemic entrenchment ranking.
We only need to show that 3.1 holds for this ranking.
16 CHAPTER 3. CONSTRUCTING CONTRACTION AND REVISION
If , A then B K : A < A B = B K : A / K A (A B) = K A and
(A B) K A. We will show that if B K then (A / K A and (A B) K A)
(A / KA and B KA). B K (KA)+A (using (K

5)) implies AB KA.


Thus if AB KA then (AB) (AB) = B KA. If B KA then obviously
(A B) K A.
If A then A K A. That is (K A) + A = K A. Using (K

5) we have
K (K A) +A = (K A). Along with (K

2) we get (K A) = K.
For the second part we can dene the function as in equation 3.1. We only need to show
that this function is a contraction.
(K

1) Suppose K A for some L. By compactness of there exist


1
,
2
, ..,
n
such that (
1

2
..
n
) . To show K A we have to show that K
and either A < A or A. Suppose , A, then A
i
> A for all
i
. Hence we have
A(
1

2
..
n
) > A as by (EE2) A(
1

2
..
n
) = (A
1
) .. (A
n
)
which is equal to some A
k
(because of (EE3)) and for which A
k
> A. Now
(EE2) also gives A (
1

2
..
n
) A which with (EE1) gives A < A .
So for n 1 above proof works. If n = 0 then hence A and by (EE2),
C A for all C L. , A and (EE5) give C > A for some C. (EE1) now gives
A > A.
(K

2) Follows from the denition.


(K

3) As A / K, K ,= K

. Hence (EE4) gives for all B L, B A. Now if K then


(EE4) gives a C such that , C or > C A. Using (EE1) and (EE2) we have
A > A or A > A.
(K

4) A A hence A = A which means that A A > A is not possible.


(K

5) Let K, using inverse of DT we have to show that A K A. By the


equation 3.1 we need to prove that either A < A(A) or A. But K so does
A . Let us assume that , A and show A < A (A ). Since A (A ),
by (EE2) we have C A (A ) for all C L. A not being a tautology satises
A < C for some C by (EE5). Thus by (EE1) we get A < A (A ).
(K

6) If A B then by (EE2) we have A = B and AC = BC. Thus A < AC


B < B C. The equality of K A and K B now follows from the denition.
(K

7) We have to show that K A K B K (A B). Let K A and K B


i.e. A < A and B < B . Thus by (EE2), (EE1) we get A B < A and
AB < B. Now (EE3) gives that (A) (B) = (A) or (B). Using
(EE1), A B < (A ) (B ) = (A B) i.e. K A B.
(K

8) A / K (A B) implies A B (A B) A = A. Now if K (A B) then


A B < (A B) . Thus by (EE1) we get (A B) > A. Finally using (EE2)
we get A (A ) (B ) > A, hence A > A and we have K A.
3.2 Partial Meet Contraction
Partial meet contraction is very intuitive way of dening contraction functions. At the rst
step a contraction KA with a reasonable A must not contain A. So we can start thinking
3.2. PARTIAL MEET CONTRACTION 17
of beliefs which dont entail A or maximal subset m of K which fail to imply A. As a
formalization think of these subsets as m satisfying following three properties-
(i) m K
(ii) A / Cn(m)
(iii) For any m

such that m m

K, A Cn(m

)
The denition gives very much what we desire, like m is closed and addition of anything
in K m to m gives K(proofs later). Surprisingly these subsets also satisfy K

(1 6),
[12, pp 6263] calls them maxichoice contractions. Here we can implement our basic rule
that logical inferences alone are insucient for structuring the belief change i.e. there has
to be a preference some where. We will see how further restrictions lead us to contraction
functions.
Let us denote all maximal subsets of K which fail to imply A in K by KA and
M(K) =

AK
KA. Note that if A / K then we take KA = K. For contraction
KA, instead of some m in KA we will choose some preferred elements of KA and take
their intersection. Let S denotes the selection function for choosing the preferred elements
in KA. Following denition may lead us to our contraction function.
K A =

S(KA) if , A
K otherwise
(3.3)
where, S(KA) KA
If function above is a contraction then it is called Partial Meet Contraction. Dierent
denitions of selection function S can lead us to dierent functions. But we will see that
the most appropriate S is when it chooses the top elements in KA based on a reexive
and transitive preference relation on all maximal sets M(K) of K.
S(KA) = m KA : m m

, m

KA (3.4)
Following results will help us in proving the theorem 5.3.1 which tells that partial meet
contraction function dened above is indeed a contraction.
Lemma 3.2.1 Let m is in KA for some A in K, then m is closed and Cn(m) = K
where K m.
Proof Let m and assume / m. Now m implies K as m K and K is
closed. Hence m m K, which implies m A or m A. But then
m A (as m ) gives contradiction.
For this we already have Cn(m) K and only K Cn(m ) has to be proved.
Let K m, we show that A m and hence any set containing A and m(in
particular Cn(m )) will imply . By contradiction suppose A / m which means
m m A K as (A ) and K. Thus (A ) A m. But
((A ) A) A gives that A m, a contradiction.
Second part of above result is very crucial. It says that all the beliefs in K outside m
behave similarly w.r.t. m. Suppose m is in KA for some A, and B K m then addition
of some in K m to m will give all of K and in particular B. Thus one can check that m
will also satisfy all the three conditions for the membership of KB. Here is the formalized
result-
18 CHAPTER 3. CONSTRUCTING CONTRACTION AND REVISION
Lemma 3.2.2 Let A K and m KA then B K m, m KB.
Using this result we compare K(A B), KA and KB. One can easily see that if
m K(A B) then it is in KA or KB because the second condition A B / m
implies A / m or B / m. What if m is in KA alone? The rst and second condition
for membership of KAB are satised. Now our previous result implies third condition.
Suppose m m

K, then in addition to A Cn(m

), Cn(m

) = K is also there implying


that B Cn(m

) (as B K). Hence A B is in Cn(m

). We present this result and then


move on to our main theorem.
Lemma 3.2.3 Let A and B are in K then K(A B)KA KB.
Theorem 3.2.1 Let be a reexive and transitive relation on M(K). Then function
dened through 3.3 with S as in 3.4, satises (K

1)-(K

8).
Proof
(K

1) This holds because each maximal subset m in S(KA) is closed and intersection
of closed sets is closed.
(K

2) Follows from the denition.


(K

3) Follows from the denition.


(K

4) Suppose A K A. This can only happen only under the alternate denition of
3.3 when K = K A and A.
(K

5) Let K. Assume , A so that rst denition of 3.3 is used, we show that


(A ) K A. Using second part of the lemma 3.2.1 with = A we get
m KA, Cn(m A) = K. Thus A m, K implying that A
K A =

S(KA). Hence (K A) +A.


(K

6) Use, A B implies KA = KB.


(K

7) If A or B then obviously this holds, if A then KA = K and KB = K(A


B)( A B B). So let us assume , A and , B. We have to show (

S(KA))
(

S(KB)) S(K(A B)) . But S(KA KB) = S(KA) S(KB)


S(KA) S(KB). Hence

(S(KAKB))

(S(KA) S(KB)). Now by


property of sets

(S(KA)S(KB)) = (

S(KA))(

S(KB)) and by lemma


5.3.3 we get (

S(KA)) (

S(KB)) =

(S(KA) S(KB))

(S(KA
KB)) =

S(KA B).
(K

8) Let A / K (A B). Thus there exists an m

K(A B) such that m


m, m K(A B) and A / m

. Note that m

KA as A / m

. Now we
show

S(K(AB))

S(KA) by showing S(KA) S(K(AB)). If m


S(KA) then m

m, m KAB implying that m

m, m K(AB)
due to transitivity of . Hence m

S(K(AB)) giving S(KA) S(K(AB)).

In the next section we present another approach for constructing contraction function.
3.3. SYSTEM OF SPHERES 19
3.3 System of Spheres
This method was employed by grove [9]. All the notations are borrowed from [8, section
4.5]. In this method we consider maximal consistent sets which are same as models as our
basic units. A set M L is called maximal consistent if -
(i) M is consistent i.e. B L s.t. B and B M.
(ii) If / M then M is inconsistent.
Let for a set of beliefs K, [K] denotes all maximal consistent extensions of K. Where a
set of beliefs M being maximal means . We use / to denote all maximal consistent sets.
Thus-
[K] = M / : K M (3.5)
For dening a contraction K A of K w.r.t. A, a system of spheres o is considered. Each
element of o is a subset of / i.e. it contains some maximal consistent sets. In addition o
satises following properties-
(i) o is totally ordered by . That is for any S, S

o either S S

or S

S.
(ii) [K] is the minimum of o i.e. if S o then [K] S.
(iii) / o.
(iv) If A is a belief and there is any sphere in o intersecting [A] then there is a smallest
sphere in S intersecting [A]. We will use S
A
for this smallest sphere.
Given the denition of o we can dene K A as those formulae of K that are near to
consequences of A. Formally-
K A =

([K] C(A)), if , A
K otherwise
where C(B) = [B] S
B
(3.6)
Simplifying the denition we get-
K A = (

[K]) (

C(A)) = K (

C(A)) (3.7)
as K =

[K]. We can see using Harper-Identity that

C(A) is nothing but revision of
K w.r.t. A. Before moving on to our main result we prove some small results.
Lemma 3.3.1 If A B then [A] [B] and S
B
S
A
.
Proof Let A B then if M [A] then B M implying that M [B]. Also S
A
[A] is
non empty implies S
A
[B] is not empty which further implies S
B
S
A
by virtue of (iv)
condition on o.
Lemma 3.3.2 [A B] = [A] [B]
Proof A, B (A B) implies [A] [B] [A B] using the previous result. For the
other way we will use following property of maximal consistent sets -
A B M then either A M or B M
If M [A B] then either A M or B M, implying M [A] or M [B] i.e.
M [A] [B].
20 CHAPTER 3. CONSTRUCTING CONTRACTION AND REVISION
Theorem 3.3.1 Let K-A be as in 3.7, then it satises K

(1 8).
Proof
(K

1) Use closeness of maximal consistent sets.


(K 2) Follows from the denition.
(K

3) Suppose A / K, thus KA is consistent (if not then KA = K

implying
A K or A K). Thus there is a maximal consistent extension of KA which
is in [K] and [A] hence S
[A]
= [K], implying K

(C(A)). Thus K A = K.
(K

4) Let , A. Thus [A] is not empty. Hence A / C(A), implying A / K A.


(K

5) Let K, we prove that A K A. Now K implies A K.


Also if M [A] then A M as A (A ) and A M. Thus
A C(A) implying A K A.
(K

6) Use, A B implies [A][B].


(K

7) We have to show K (

C(A)) (

C(B)) = K (

(C(A) C(B)))
K (

(C((A B)))). We will show that C((A B)) C(A) C(B). So,
C((AB)) = S
(AB)
[(AB)] = S
AB
[AB] = S
AB
([A][B]) =
(S
AB
[A]) (S
AB
[B]). But then A, B (A B) and previous
results imply (S
AB
[A]) (S
AB
[B]) (S
A
[A]) (S
B
[B]) =
C(A) C(B).
(K

8) Let A / K(AB) = K(

C(AB)), we have to show K(

C(AB))
K (

C(A)). Assume A K, then there is an M C((A B)) = S


(AB)

[(A B)] such that A / M. But then M [(A B)] = [A] [B] and A / M
implies M [A]. Thus S
(AB)
[A] is non-empty, implying S
A
S
(AB)
. Also
A (AB) implies S
AB
S
A
. Thus we get S
A
= S
(AB)
which along with
[A] [(AB)] gives C(A) = S
A
[A] = S
(AB)
[A] S
(AB)
[(AB)] =
C((A B)). Hence we get

C((A B))

C(A).
3.4 Implementation of Belief Revision
This section describes how a computer implementation of a contraction or revision func-
tion can be achieved. With every implementation there is always an overhead of proving
theorems eciently. Even if one ignores that, from computational point of view, it seems
impossible to deal directly with belief sets. For implementation, methods of partial meet
contraction and spheres are completely out of question. These methods use closed sets
as their basic units which are simply too big. Epistemic entrenchment (EE) seems to be
plausible choice because of explicit denition of a contraction. But then diculty arise as
ordering requires ranks on typically innite number of sentences. Also EE ordering is lost
in process of change, and as a consequence the iteration of change functions is not naturally
supported. In this chapter we describe a computational model which uses nite partial
entrenchment(FPE) ranking [1, chapter 15], [13]. This ranking is essentially a nite repre-
sentation of EE ranking. After describing FPE we give a PROLOG based implementation
of contraction functions.
3.4. IMPLEMENTATION OF BELIEF REVISION 21
3.4.1 Finite Partial Entrenchment Ranking
FPE grades the content of a nite knowledge base according to its epistemic importance.
Formally this ranking maps a nite set of sentences to rational numbers. The higher the
value assigned to a sentence the more rmly held it is, or the more entrenched it is.
Denition 3.4.1 A FPE ranking is a function B from a nite subset of sentences into
the interval [0, 1] such that the following conditions are satised for all domain(B) =
dom(B):
(i) dom(B) : B() < B() , if is not a tautology
(ii) If , then B() = 0
(iii) B() = 1 if and only if
Here 1
st
and 2
nd
conditions only put restrictions on tautologies and contradictory propo-
sitions. The rst condition says that a sentence assigned higher value than an arbitrary
sentence , do not entail . The sentences mapped to numbers greater than zero represent
the explicit beliefs, and their logical closure represents its implicit beliefs-
Denition 3.4.2 Let B denotes family of all FPE rankings. Then for some B in B,
exp(B) = dom(B) : B() > 0 and content(B) = Cn(exp(B)).
Let be a nontautological sentence and B be a nite partial entrenchment ranking then
degree of acceptance of is-
degree(B, ) =

largest j such that exp(B) : B() j if content(B)


0 otherwise
Theorem below, tells how a FPE ranking can generate an EE ordering using degree of
acceptance. The proof is straight forward.
Theorem 3.4.1 Let B B and , be propositions. Dene
B
by
B
i , or
degree(B, ) degree(B, ). Then
B
is an EE ordering related to content(B).
Thus we have solved the problem of dening ordering on an innite set by dening it on
its nite representation. But how one can design the contraction or revision using all this is
yet be seen. Actually instead of only dening contractions we will setup a general function
which revises the degree of given proposition . So by changing degree to 0 we can employ
contraction and in general we can move sentences up or down as the preference changes.
For this adjustments are dened, which take a FPE B a proposition , a rational in [0, 1]
and give a new FPE in which degree() is i.
Denition 3.4.3 The adjustment of a nite partial entrenchment ranking B is a function
such that
B (, i) =

(, i) if i degree(B, )
(B

(, 0))
+
(, i) otherwise
where
B

(, i)() =

i if degree(B, ) = degree(B, ) and B() > i


B() otherwise
for all dom(B)
B
+
(, i)() =

B() if B() > i


i if orB() i < degree(B, )
degree(B, ) otherwise
22 CHAPTER 3. CONSTRUCTING CONTRACTION AND REVISION
for all dom(B)
With this denition we try to implement adjustment function in PROLOG.
3.4.2 A PROLOG Implementation
For calculating degree we use the function below
FUNCTION degree(B, a)
degree(B,a) max(B)+1
if dom(B) then bottom min(B)
else bottom min(B)-1
do degree(B, a) degree(B,a)-1
until degree(B,a) = bottom or b dom(B):B(b) degree(B,a) a
Then the adjustment function is dened as in the denition. We achieve negation ()
of a proposition in PROLOG by failure. Function not(X) fails if X is success
not(not(X)):- X.
not(X):- X,!,fail;true
For OR and AND ; and, are used respectively. Implication is achieved by the denition
symbol : , suppose we want to include a b as our belief then b : a is loaded in the
database. So whenever b is searched for success and it reaches this proposition, the success
of a is checked. If a is a success then b also becomes success. For adjustment function we
consulted function denition given on page 204 of [1]. Here is the PROLOG code for main
functions
adjust(B,A,I):-
length(B,N1),Len1 is N1
*
1.0,
enterAll(B),
removeAll(B),
findplace(B,Len1,A,I,Ranki,B1),
length(B1,N),Len is N
*
1.0,
rank(B1,A,Ranka),
newB(B1,Ranka
*
2,Degreea),
(Degreea>=I,!,
movedown(B1,A,Ranka,Ranki,NewB)
;
moveup(B1,A,Ranka,Ranki,NewB)
),
write(NewB).
movedown(B,A,Ranka,Ranki,NewB):-
removeA(B,Ranka
*
2-1,A,NewB1),
addA(NewB1,Ranki
*
2-1,A,NewB2),
forD(B,NewB2,0.0,A,Ranka,Ranki,NewB).
moveup(B,A,Ranka,Ranki,NewB):-
rank(B,not(A),RankNa),
lengthExp(B,Lenexp),
length(B,Len),
(RankNa=<Lenexp/2,!,
movedown(B,not(A),RankNa,Len/2,NewB1)
;
copy(B,NewB1)
),
3.4. IMPLEMENTATION OF BELIEF REVISION 23
removeA(NewB1,Ranka
*
2-1,A,NewB2),
addA(NewB2,Ranki
*
2-1,A,NewB3),
forU(NewB1,Lenexp/2,NewB3,Ranki+1,A,Ranka,Ranki,NewB).
findplace(B,Len,A,I,Ranki,NewB):-
Iter is 1.0,
while1(B,Len,Iter,Ranki1,I),
newB(B,(Ranki1)
*
2,Btwoi),
(I>Btwoi,!,
insert(B,Ranki1
*
2-2,[],NewB1),
insert(NewB1,Ranki1
*
2-1,I,NewB),
Ranki is Ranki1
;
Ranki is Ranki1,
copy(B,NewB)
).
rank(B,A,Rank):-
length(B,Len), % Actual length of B.
lengthExp(B,Lenexp),% length corresponding to explicit part of B.
enterExp(B), % enter explicit statements into the prolog database.
(A,!,
prove(B,A,Lenexp/2,Rank)
;
Rank is Len/2,
removeExp(B) % remove explicit statements of B.
).
The function with the argument look like this
adjust([[(ki:-not(fi))],0.4,[(si:-fi)],0.3,[fi],0.25,[ki],0.1,[],
0.0]],fi,0.6).
The rst argument is a list and represents the FPE which has to be adjusted, 2i1
th
2i
th
element starting from index 1 represent the belief and its ranking respectively. The second
argument is the belief which has to be revised and third argument is the new place for the
revised belief(in this case fi). The results obtained are not outstanding, the implementation
of as failure is not sound. So some times some beliefs are retracted when they shouldnt.
Limited programming structures and implementation of everything as a recursion makes
situation more complex.
Chapter 4
Three-Valued Logic and Belief
Revision
Three-valued logic or Lukaseiwicz logic consists of a third truth-value 1/2 between 0 (false-
hood) and 1(truth). The interpretation of p = 1/2 is that p is undecidable in current
context. Like sentences of the type It will rain tomorrow are not decidable and can be
given value 1/2. This possibility motivated us to look at three-valued logic from belief
change point of view. We want our belief sets to be more vague, where there is greater
variety of beliefs and not just this happens or that doesnt happen.
4.1 The Logic
The propositions are as usual ws formed using the connectives , , with xy (x
y) y and x y (x y). The interpretation of sentences is in L
3
= 0, 1/2, 1,
using-
v(x) = 1 v(x)
v(x y) = min(1, 1 v(x) +v(y))
(4.1)
Using the equivalent denitions, interpretation rule for and are -
v(x y) = max(v(x), v(y))
v(x y) = min(v(x), v(y))
(4.2)
Besides this we dene two more connectives ,
A (A A)&(A B) (A (A B))
The rst axiomatization of three-valued logic was given by Wajsberg in 1931 ([5]) using
following axioms
A1 p (q p)
A2 (p q) ((q r) (p r))
A3 ((p p) p) p
A4 (p q) (q p)
with modus ponens as rule of inference:
p, p q
q
The three-valued propositional calculus on L obtained by above axioms and rule also satisfy
following properties
24
4.2. A SEQUENT CALCULUS 25
(1) p q p
(2) p q, q r (p r)
(3) p (p q) q
(4) p p
(5) p p
(6) p (p q)
(7) p p
(8) p p
along with the deduction theorem
If then ( ) or
Similar to classical logic, three-valued logic is compact. The denition of consistency is
same i.e. a consistent set mustnt entail a proposition and its negation (). The maximal
consistent sets are consistent and inclusion of any formula outside it makes it inconsistent.
Similar to classical logic, every consistent set has a maximal consistent extension([5, page
467]). Finally the usual completeness and soundness results hold here.
During the study of three-valued logic we thought of derivability in detail. We also looked
for other possible axiomatizations. In the end we were able to derive a sequent calculus
for this logic. Following sections will presents the sequent calculus. For proving theorems
in this logic in general we generalized automated deduction methods for the classical case
and implemented them. Many theorems in later sections are proved using these automated
methods. In the following sections we will also present the automated proof methods for
three-valued logic.
4.2 A Sequent Calculus
In [10, Section 3.3] gentzen sequents were used for presenting a sequent calculus. For this
logic we consider sequents of a dierent type. Sequents will be of the form
;


with ,

, nite. Premises(assumptions) are of two types i.e. formulae in are dierent


from those in

. One can think of

as of lesser credibility than . The set of conclusions


however has only one status. Below we present the rule for our sequent calculus.
Axiom Rule:
, ;

, ,
Logical Rules:
, ,


(L)
;

, ,

,
;

,
(R)
;

, ,

, , ;

,
;


(M)
;

, ;

,
(L)
;

,
;

,
(R)
26 CHAPTER 4. THREE-VALUED LOGIC AND BELIEF REVISION
;

, ,

, ;

, ,


(M)
;

,
,


( L)
, ;

, , ;

,
;

,
( R)
;

, , ;

, ,

,
;


( M)
;

,
, ;


(L)
;

,
;

,
(M)
, ;

,
(R)
4.2.1 Truth of a Sequent
Having dened the sequents we need to give semantics associated with them. Let ;


with ,

& nite, be a sequent and v be a valuation. Then ;

is said to be true
under v and written v [= ;

if and only if there is a w ,

or such that
v() = 0, 1/2 or 1 respectively. A valid sequent is true under all valuations and is denoted
as [= ;

.
4.2.2 Provability of a Sequent
A sequent ;

is provable if there exists a derivation where ;

is root and
instances of Axiom Rule are the leaves. Root is obtained by gradual application of sequent
calculus rules on leaves. A provable sequent is written as ;

.
4.2.3 Equivalence
Soundness:If ;

then [= ;

.
Proof Since the sequent is provable, the derivation using the rules is nite and all the
branches start from axiom rule. Now any valuation v can assign only three values 0,1/2,1
to the repeating formula in the instance of axiom rule and hence according to our def-
inition the sequent is true under all valuations. Now all other sequent calculus rules are
constructed in a manner that, premise sequent is true under a valuation if and only if the
conclusion is true under it. Thus using the derivation we can conclude that our provable
sequent is also true under all valuations.
Completeness:If [= ;

then ;

.
Proof Let S be the set of sequents s.t.

= and f be the strategy used for deriva-


tion i.e. f takes a sequent in S to a non-proposition in

. Now given the sequent


;

build the derivation tree T whose root is this sequent and nodes are formed
using the rules in sequent calculus. Formally if ;

/ S then it is a leaf otherwise


children of ;

are the premises of the rule corresponding to = f(;

).
Since ,

and are nite T is nite. So we only need to prove that leaves of T are
nothing but instances of axiom rule. On the contrary let us assume that there is a leaf

1
;

1

1
such that
1

1

1
= . As rules are dened for all possible occur-
rences of a given type of formula this leaf can only have propositional letters. Now all
4.3. AUTOMATED PROOF METHODS 27
our rules are dened in a way that premise are valid if and only if conclusion is valid
and we have that
1
;

1

1
is valid as ;

is valid. But a valuation v s.t.


v(a) = 1 if a /
1
, v(a) = 1/2 if a /

1
and v(a) = 0 if a /
1
can make the sequent false,
leading to a contradiction. Hence leaves are instances of axiom rule and proof is nished.

4.3 Automated Proof Methods


The automated proof methods solve the following problem. Given a w in L with the
corresponding derivability relation , is ? In [10, chapter 2] various automated methods
for classical logic are considered. For three valued logic we generalize the sematic tableaux
method and convert binary decision diagrams (BDD) to tertiary decision diagrams (TDD).
4.3.1 Semantic Tableaux
A semantic tableaux method can be devised on the basis of above sequent calculus. Sequents
of the form ; are same as theorems in this logic and hence checking validity or theorem
proving can be easily automated. Following notations are used in production of the tableaux

+ := all valuations v s.t. v() = 1.


:= all valuations v s.t. v() = 1/2.
:= all valuations v s.t. v() = 0.
Suppose we want to check for validity of a w . We want to see if the set of valuations
making 1 is indeed the set of all valuations. So we start with + and form a tableaux of
signed w using following rules-
For
+
+
For
+( ) ( ) ( )
+
+
If there is a branching in the rules then it means that valuations which satisfy the
predecessor formula are the valuations which satisfy all the branches. Valuations satisfying
a branch corresponds to union of the all valuations of formula in the branch. Finally if
there is no scope for further application of any of the rules, tableaux has to be checked for
closeness. A branch is said to be closed if there is a variable a such that all of +a, a and
a are present in it and a sequent is valid if all its branches are closed.
28 CHAPTER 4. THREE-VALUED LOGIC AND BELIEF REVISION
For
+( ) ( ) ( )
+ + + +

For
+( ) ( ) ( )
+ +
+ + +

4.3.2 Tertiary Decision Diagrams
Tertiary Decision diagram of a w is a tree in which each vertex is a propositional variable
and has three down going edges each for three possible valuations. We redene the methods
A
T
T
F
B
U
T
A=F
A=U
A=T
B=T
B=U
B=F
Figure 4.1: TDD of (p q), U is for truth value 1/2
in [10, Chapter 2, page 63] to get TDD. We have a recursive function Convert(p) which
takes a w p and returns its TDD. Convert(p) uses four other routines
(1) TDDneg(TDDp)- If TDDp is TDD of p then this function returns TDD of p
(2) TDDor(TDDp,TDDq) - This returns TDD of p q
(3) TDDand(TDDp,TDDq) - This return TDD of p q
(4) TDDimp(TDDp,TDDq) - This returns TDD of p q
(5) TDDmake(p, TDD1, TDD2, TDD3) - This function returns a TDD whose top node is
p and other three TDDs are its children. This function additionally checks that all
the three TDDs are not equal, if they come out equal then it returns one of the TDDs
discarding p.
Each of the rst four routines use TDDmake to join two trees. For example TDDor
does the following:
TDDor(T,) = T, TDDor(F, ) TDD, TDDor(U,U)=U
4.4. CONTRACTION AND REVISION 29
If = A
+
;

and

= A

+
;

then
- if A < A

then
TDD(,

) = TDDmake(A, TDDor(
+
,

),TDDor(

),TDDor(

))
- if A > A

then
TDD(,

) = TDDmake(A, TDDor(,

+
),TDDor(,

),TDDor(,

))
- if A = A

then
TDD(,

) = TDDmake(A, TDDor(
+
,

+
),TDDor(

),TDDor(

))
4.4 Contraction and Revision
In this logic we expect some modications in contraction and revision postulates. First of
all we note that here A, A as well as A, A are inconsistent. So removal of A alone,
doesnt ensure safety when A has to be added. We have to remove A from our belief set
K before adding A. Thus Levi-Identity has to be changed. Also in Harper-identity we see
that K A and K A both dont have A implying that their intersection with K may
yield a contraction. But then due to p p it is expected that K A is farther than
K A w.r.t. consequences of A. Hence KK A is probably smaller than KK A,
forcing us to change Harper-Identity too. With the changed denitions and establishing
interrelationships we reach following denitions for revision and contraction.
Postulates for Contraction
(K

1) K A is a belief set
(K

2) K A K
(K

3) If A / K then K A = K
(K

4) If ,A then A / K A
(K

5) K (K A) +A
(K

6) If A B then K A = K B
(K

7) K A K B K (A B)
(K

8) If A / K (A B) then K (A B) K A
Postulates for Revision
(K

1) K A is a belief set
(K

2) A K A
(K

3) K A K +A
(K

4) If A / K then K +A K A
(K

5) K A = K

if and only if A
(K

6) If A B then K A = K B
(K

7) K (A B) (K A) +B
(K

8) If B / K A then (K A) +B K (A B)
Now we discuss the reasons behind changing postulates. For Contraction we only
changed (K

6). We realized while proving the Harper-Identity that A B does not


imply A B, but A B is implied. But since that is same as A B we
make the change. For similar reasons we changed K

6. For K

4 we see that version with


A implies our postulate. So we preferably used the weaker postulate. K

5 is straight from
the following property of three valued logic, if K A is inconsistent then A K.
30 CHAPTER 4. THREE-VALUED LOGIC AND BELIEF REVISION
4.5 Interrelations
Changes in epistemic states modelled as expansion, revision and contraction are interrelated.
Revision of a belief set K with respect to a sentence A can be seen as removal of any
incompatible information A and then addition of A along with all logical consequences.
In 3valued logic the Levi-identity takes the following form-
K A = (K A) +A (4.3)
Let us call the revision function above R(). Following theorem tells that function R()
dened through Levi - identity is indeed a revision function.
Theorem 4.5.1 If a contraction function satises (K

1) to (K

6) then R() satises


(K

1) to (K

6). Furthermore if also satises (K

7) and (K

8) then R() satises


(K

7) and (K

8) respectively.
Proof
(K

1) From denition.
(K

2) A A hence (K A) +A A.
(K

3) From (K

2), K A K and hence (K A) +A K +A.


(K

4) Let A / K. Using (K

3), K A = K and hence K A+A = K +A.


(K

5) if Let A. Thus A, A (KA) +A. But then K A = (KA) +A = K

as B L, A, A B.
only if Let (K A) + A = K

. Thus in particular (K A) + A A. Using


D.T. we get K A (A A) which with the theorem (A A) A gives
K A A.
(K

6) Use A B if and only if A B and (K

6).
(K

7) Let K A B = K (A B) + (A B). Thus (A B) K (A


B) = K (A B) as De Morgan holds for , i.e. (A B) (A B).
Also (A B) (A B) A. So we will show that (A B)
K (A B) which by (K

7) will give (A B) K (A B) K
(A B) K ((A B) (A B)) = K (A). From where we can
get (K A+A B) = (K A+A) +B.
(A B) K (A B) can be shown by using (K

5), (A B)
K (A B) K K (A B) + (A B). Applying (D.T.) ((A
B) ((A B) )) K (A B). But ((A B) ((A B)
)) ((A B) ). Hence we get (A B) K (A B).
(K

8) Assume B / K A +A and let (K A +A) +B. From the assumption


we get (A B) / K A. Also (A B ) K A = K (A B) A
because A ((A B)A). Using (K

8) with A

= (A B) and B

= A
we get K A K (A B). Hence we get K (A B) + (A B) and
our proof is complete as (A B) (A B).
4.6. EPISTEMIC ENTRENCHMENT 31
Similar to Levi identity, Harper Identity takes the following form-
K A = K K (A) (4.4)
Let us call revision function dened above C().
Theorem 4.5.2 If a revision function * satises (K

1) to (K

6) then C(*) satises


(K

1) to (K

6). Furthermore, if * also satises (K

7) and (K

8) then C(*) satises


(K

7) and (K

8) respectively.
Proof
(K

1) Intersection of two belief sets is also a belief set.


(K

2) Follows from denition.


(K

3) A / K implies A / K as A A. Using (K

4) we get K A = K +A
which means K K A = K.
(K

4) Let A KA = KKA. Thus both A and A KA giving KA = K

.
But then (K

5) gives A.
(K

5) Let K, we need to show (K K A) + A. (A ) implies


A K which implies (A ) K. So our proof will be complete after showing
(A ) K A. But that is ne as A (A ) and A K A because
of (K

2).
(K

6) Use A B i A B and (K

6).
(K

7) Following results will be assumed during the proof:


(i) (A B) (A B)
(ii) ((A B) A) A.
Now K A K B means K, K A, K B. Taking A

= (A B)
and B

= (A) and using (K

7) we get K ((A B) A) = K A
K (A B) +A. Similarly we can get K B K (A B) +B. Using
D.T., A , B K (A B). Now C() already satises (K

5) so
K (K K (A B)) +A B. And we have (A B ) K (A B).
But (A B ), (A ), (B ) because = 0 or 1/2 demands
that one of the premises is not equal to 1 and hence K (A B). Thus
K K (A B) = K (A B).
(K

8) Let A / K (A B). Which means A / K or A / K (A B). If A / K then


K(AB) K = KA. So consider the case when A K but A / K (AB).
Taking A

= A B, B =

A and using (K

8) we have K (A B) + A
K (AB) A = K A. But K (AB) K (AB) +A and we have
K (A B) = K K (A B) K K A = K A.
4.6 Epistemic Entrenchment
In 3-valued logic the epistemic entrenchment doesnt change signicantly. Epistemic en-
trenchment is mostly independent of nature of the inference relation. It only sees it as a
preference relation over the language. We will see that the earlier denition of contraction
in terms of Epistemic entrenchment works here.
32 CHAPTER 4. THREE-VALUED LOGIC AND BELIEF REVISION
Theorem 4.6.1 Let be a contraction function and let K is a belief set. Then the ordering
dened in 3.2 is an epistemic entrenchment ranking.
Proof Follows line by line similar to the classical case.
Theorem 4.6.2 Let K be a belief set. For every contraction function of K there is an epis-
temic entrenchment (related to K) such that 3.1 is true. Conversely for every epistemic
entrenchment related to K, there is a contraction function satisfying 3.1.
Proof For the rst part theorem 4.6.1 gives the required epistemic entrenchment ranking.
We only need to show that 3.1 holds for this ranking.
If , A then B K : A < A B = B K : A / K A (A B) = K A
and (A B) K A. We will show that if B K then the constraint (A / K A
and (A B) K A) on the above set is same as (A / K A and B K A) which
is same as K A.If B K A then obviously (A B) K A. For the other way
(B B) and B K implies B K and K A = K A. Using (K

5) as
K (KA)+A implies (A B) KA. Now (AB) (AB)
implies (A B) K A. But then (A B), (A B) B as for
B = 0, 1/2 B = 0 implying that both the premises cant be 1 and hence B K A.
If A then A K A. That is (K A) + A = K A. Using (K

5) we have
K (K A) +A = (K A). Along with (K

2) we get (K A) = K.
For the second part we can dene the function as in equation 3.1. We only need to show
that this function is a contraction. All the postulates go through easily except the (K

5).
Below is the proof-
(K

5) Let K, we will show that (A ) K A. By the equation 3.1 we need to


prove that either A < A (A ) or A. Note that K so is (A ). Let us
assume that , A and show A < A (A ). Since A (A ), by (EE2) we
have C A (A ) for all C L. A not being a tautology satises A < C for
some C by (EE5). Thus by (EE1) we get A < A (A ).
4.7 Partial Meet Contraction
As in the classical case this method also leads us to contraction functions. Retraction
of a belief can be modelled through maximal subsets of K which fail to imply K. From
practical view point that is very much expected because addition of beliefs with possible
connotation wont make any dierences as far as belief change is concerned. With the
assumed background we move on to proving theorem similar to 5.3.1. But before that we
make a small comment on why the lemma 3.2.1 should hold here. In the proof of lemma
3.2.1 we make use of modus ponens just after deduction theorem, here similar result can
be obtained with on which modus ponens works. Also in the second part we use modus
ponens on ((A ) A) A, which works here because following weaker result holds
((A ) A) A.
Theorem 4.7.1 Let be a reexive and transitive relation on M(K). Then function
dened through equation 3.3 with S as in equation 3.4, satises (K

1)-(K

8).
Proof
4.8. SYSTEM OF SPHERES 33
(K

1) We show that each m KA is closed i.e. it is a belief set. And we already know
that intersection of closed sets is closed. Suppose m now suppose / m thus
m m K. Hence m A or m ( A), but then m A as m
giving contradiction.
(K

2) Follows from the denition.


(K

3) Follows from the denition.


(K

4) Suppose A K A. This can only happen when K = K A and A.


(K

5) Let K. We show that (A ) K A, suppose not then there is a m


S(KA) such that (A ) / m. Thus m m (A ) K. Which implies
((A ) A) m, but then ((A ) A) A further implies A m leading
to contradiction. So (A ) K A and hence K A+A.
(K

6) Use, A B implies KA = KB.


(K

7) This proof goes similar to the classical case as in proof of theorem because (A
B) B holds here.
(K

8) Here too we dont use anything new, hence classical proof goes through.
4.8 System of Spheres
System of spheres derive contraction by maximal consistent sets. Since three-valued logic
distinguishes between A and A we have to be careful of denition of consistent sets. We
see that p p but converse is not true. If we dene consistency of K L as
B L such that B and B K
then there is no dierence. One can prove that this denition is equivalent to classical
denition of consistency. If we look at maximal consistent set M then also they satisfy
most properties of classical maximal consistent sets. Like if M then M. Hence we
dont need to alter the denition of maximal consistent sets. We need to change denition
of contraction. Since p, p is inconsistent, its sucient to take beliefs in KA which are
near to consequences of A. The following modication as expected serves the purpose-
K A = (

[K]) (

C(A)) = K (

C(A)) (4.5)
Before proving that above function is a contraction, we prove that maximal consistent sets
satisfy the property-
A B M then either A M or B M
Proof Let A B M but A, B / M. Hence MA and MB are inconsistent. Thus
A B M. But (A B) (A B) implies (A B) M implying M is
inconsistent, a contradiction. Besides this results concerning S
A
and [A] also hold. One can
easily show that if A B then S
B
S
A
and [A] [B].
Here comes our nal theorem-
Theorem 4.8.1 Let K A be as in 4.5, then it satises K

(1 8).
34 CHAPTER 4. THREE-VALUED LOGIC AND BELIEF REVISION
Proof
(K

1) Since maximal consistent sets are closed here, there intersection is also closed.
(K

2) Follows from the denition.


(K

3) Suppose A / K, thus KA is consistent (if not then KA = K

, implying
A K or A K (as A A)). Thus there is a maximal consistent extension
of K A which is in [K] and [A] hence S
[A]
= [K], implying K

(C(A)).
Thus K A = K.
(K

4) Let , A. Thus [A] is not empty. Hence A / C(A), implying A / K A.


(K

5) Let K, we prove that A K A. Now K implies A K.


Also if M [A] then A M as A (A ) and A M. Thus
A C(A) implying A K A.
(K

6) Use [A] = [B] as A is same as (A B).


(K

7) The classical proof goes through in our 3-valued context as [AB] = [A] [B] and
(A B) (A B).
(K

8) The classical proof goes through here too.


4.9 Interpretation and Applicability
In earlier sections we noticed how consistency and closure of belief sets have helped us in
modifying the theory of belief change so that it gets a sleek structure in three-valued logic.
In this section we discuss the interpretations of the obtained results. First of all we realize
that A is the central agent around which all the changes happened. In classical logic
A and A were same meaning belief in A does not happen. But here since they are not
same we have to nd new interpretations. We know that A is stronger than A. So
interpretation of A remains the same. But what about A? We guess the interpretation
of A as possibly A doesnt hold. If we assume this interpretation then the immediate
question is why A, A is inconsistent. The answer is, if A is in my belief set then i
believe that A must happen. Now in such a situation showing a positive condence in A
by means of A is clearly unacceptable. Its like ignoring ones beliefs. The existence of
possiblity notion in the form of A motivated us to look for alternative axiomatization
of 3-valued logic. In [4] and [2] we nd close relationships between S
5
which has modalities
and three-valued logic. Following theorem from [4]([2]) tells us that interpretation of A is
correct. And in general we can give possibility connotation to any belief.
Theorem 4.9.1 A Wajsberg algebra < X, , , 1 > is equivalent to a 3-valued Lukasiewicz
algebra < A, , , , , M, 0, 1 > via the transformation.
a b (Ma b) (Mb a)
a b (a b) b
a B (a b)
Ma a a
0 1
4.10. GENERALIZATION TO N-VALUED LOGICS 35
This theorem is interesting because algebra is in terms of a new connective M. In this alge-
bra our A takes the form MA which means possibly A does not happen, if interpretation
of M is taken as it is possible that. Our next theorem which is borrowed from [2] ensures
that M behaves exactly like possibility connective.
Theorem 4.9.2 3-valued Lukasiewicz logic is embeddable into S
5
.
Embeddable means for every w in 3-valued logic there is a translation

of in
S
5
such that is a theorem in 3-valued logic if and only if

is a theorem in S
5
. The
embedding maps the connective M of Lukasiewicz algebra to possibility connective M of
S
5
. Hence our guess is perfect and our beliefs can be made vague by saying MA instead of
A.
After knowing the interpretation there are numerous uses of this kind of belief change.
In fact, since we dont loose anything as far as belief change in classical background is con-
cerned, one should think of it as fundamental. In day-today life we actually have possibility
notions in our beliefs.
4.10 Generalization to n-valued Logics
The belief change in 3-valued logics can be extended to n-valued logics. n-valued logics take
the n truth values L
n
= 0, 1/(n 1), ..., (n 2)/(n 1), 1. Interpretation of connectives
is same as that of 3-valued logic. In most of the 3-valued logic proofs, we noticed the
adjustments in classical proofs by replacing by and by . We also observed that
modus ponens with produces same results as with . So we expect our 3-valued results
to generalize for n-valued logics where we take A A (A (A ...(A

n2 times
A))..)
with A B A (A (A ...(A

n1 times
B))..). Following result will help us in embedding
classical logic into any n-valued logic. Since embedding is in terms of and , it is very
useful in generalization.
Lemma 4.10.1 Suppose is a formula in L and

is formed by replacing by and


by . Also suppose n
th
lift of a classical valuation v
2
is a valuation v
n

based on L
n
such
that v
2
(p) v
n

(p) v
2
(p) + (n 2)/(n 1), for each propositional variable - p. Then for
any ,
v
2
() v
n

) v
2
() + (n 2)/(n 1) (4.6)
Proof We proceed by induction on length [[ of . When [[ = 1 then is a propositional
variable, for which inequality holds by assumption. Suppose inequality holds for any such
that [[ m. We prove that inequality holds for [[ = m+ 1:
(i) . If v
2
() = 0 then we only need to show v
n

) (n 2)/(n 1). v
2
() =
1 v
2
() = 1 and [[ = m gives 1 v
n

) v
2
() + (n 2)/(n 1) that is
v
n

) = 1. Hence v
n

) = v
n

) = 0 (n 2)/(n 1).
If v
2
() = 1 then we show that v
n

) = 1. v
2
() = 1 1 = 0, thus 0 v
n

)
(n 2)/(n 1). Hence v
n

) = v
n

) = 1.
(ii) . If v
2
( ) = 0 then v
2
() = 1 and v
2
() = 0. Using induction we
have v
n

) = 1 and v
n

) (n 2)/(n 1). Since v


n
(b c) = v
n
(c), whenever
v
n
(b) = 1, we get v
n

) = v
n

) = v
n

) (n 2)/(n 1).
36 CHAPTER 4. THREE-VALUED LOGIC AND BELIEF REVISION
If v
2
( ) = 1 then if v
2
() = 0 then 0 v
n

) (n 2)/(n 1). But


v
n
(b c) = 1 if v
n
(b) < 1, thus v
n

) = 1. If v
2
() = 1 then v
2
() = 1 and hence
v
n

) = 1 and v
n

) = 1. Since v
n
(b) = 1 implies v
n
(b c) = v
n
(c), we have
v
n

) = v
n

) = 1.
(iii) or . Follows easily from the denitions v( ) = max(v(), v()) and
v( ) = min(v(), v()).
Theorem 4.10.1

is a theorem in n-valued logic if and only if is a theorem in classical


logic.
Proof We use completeness of any n-valued logic. Some formula is a theorem if it is 1 on
all valuations. Only if part follows immediately because any theorem in n-valued logic is
a classical theorem: any classical valuation(a valuation on L
2
) is also a valuation on L
n
,
and hence if something is true on any valuation on L
n
then it is true on any valuation on
L
2
. Thus

is a theorem in classical logic, where , are equivalent to , . Hence

and thus is a theorem.


For the if part we show that

is 1 on all valuations v
n
based on L
n
. Suppose v
n
is
some valuation based on L
n
. Then there exists a classical valuation v
2
whose n
th
lift is
v
n
, that is v
2
(p) v
n
(p) v
2
(p) + (n 2)/(n 1) for all propositional variables p. Now
using previous lemma we have v
2
() v
n
(

) (n 2/(n 1)). Since v


2
() = 1 we get
v
n
(

) = 1 and hence

is true on all valuations v


n
based on L
n
, that is to say

is a
theorem in n-valued logic.
Now all the results obtained in 3-valued logic can be generalized to n valued logic.
Thus if we do belief changes with A instead of A then we get all the belief change results
in any n-valued logic. Making comparisons with existing approaches like [7] where belnaps
logic is used might be interesting.
Chapter 5
Rough Belief Revision
In this chapter we will see how rough logic can be used to do belief change. In 3-valued
logic we were able to provide room for vagueness by realizing the notion of possibility of
a belief. There a belief A can have a weaker assertion M(A). Here in rough logic we dont
introduce new assertions of a belief, instead we make them weaker by taking incomplete
evidences as beliefs. So a belief may fail as its truth was dependent upon some incomplete
information. Most denitions and initial results are borrowed from [3].
5.1 Rough Truth
The notion of rough truth was introduced in [15] as a part of the rst formal proposal on
reasoning with rough sets. It was proposed to reect inductive truth, i.e. truth relative to
our present state of knowledge, and one that, with gain of knowledge, leads to total, deduc-
tive truth. This sense of gradualness nds an expression in, possibly, the only qualitative
version of approximate orsoft truth, as opposed to other quantitative denitions found
in, e.g., probabilistic, multi-valued or fuzzy logics.
It has generally been accepted that the propositional aspects of rough set theory are
adequately expressed by the modal system S
5
. An S
5
(Kripke) model (X, R, ) is essentially
an approximation space (X, R) (X is some set with equivalence relation R on it) where
X ,= , with the function interpreting every well-formed formula(w) of S
5
as a rough set
in (X, R). If L, M denote the necessity and possibility connectives respectively, a modal w
L(M), representing denitely (possibly) , is interpreted by as the lower (upper)
approximation () x X : all x

related to x are in (), i.e. if xRx

then x

()
(() (()
c
)
c
) of the set (). In general for a subset A of X we can dene lower(A)
and upper(A) of as:
A =

[x] : x X, [x] A
A =

[x] : x X, [x] A ,=
where [x] is equivalence class of x
Now with all the above background, a w may be termed roughly true in /
(X, R, ), if () = X. The notion of truth can be extended to rough validity which can
be further used to dene rough consequence and rough (in)consistencey. The language of
rough logic L
R
is that of a normal modal propositional logic. In the following is any set
of ws, , any ws of L
R
.
37
38 CHAPTER 5. ROUGH BELIEF REVISION
5.1.1 The semantics
Denition 5.1.1 An S
5
-model / (X, R, ) is a rough model of , if and only if every
member of is roughly true in /, i.e. () = X.
Denition 5.1.2 is a rough semantic consequence of (denoted [ ) if and only if
every rough model of is a rough model of . If is empty, is said to be roughly valid,
written [ .
We present now the logic of L
R
which will serve as the base logic for rough belief change.
As already said the language is of a normal modal logic and let
S5
denotes the derivability
relation in S5. We consider two rules of inference:
R1. R2. M
M
where
S5
M M M M
The consequence relation dening the system L
R
is given as follows.
Denition 5.1.3 is a rough consequence of (denoted
R
) if and only if there is a
sequence
1
, ...,
n
( ) such that each
i
(i = 1, ..., n) is either (i) a theorem of S
5
, or (ii)
a member of , or (iii) derived from some of
1
, ...,
i1
by R1 or R2.
If is empty, is said to be a rough theorem, written
R
.
Some derived rules of inference:
DR1. DR2. (M)L

S5
(M)L

DR3. M DR4. M
M M

DR5. M DR6.
M
DR7. M DR8.
S5

M
M M M
Some immediate results are -
Theorem 5.1.1 (Deduction) For any , , , if
R
then
R
.
Theorem 5.1.2 (Soundness & completeness)
R
if and only if [ .
Theorem 5.1.3 L
R
is paraconsistent.
We state the following without proof. One may remark though, that the proof of
Theorem 5.1.4 uses only S
5
properties, and that of Theorem 5.1.6 uses the rule of inference
R2 (in fact, DR3).
5.2. ROUGH BELIEF CHANGE 39
Theorem 5.1.4 is roughly consistent if and only if it has a rough model.
Theorem 5.1.5
R
if and only if M = M :
S5
M
Theorem 5.1.6 If is not roughly consistent then
R
for every w .
Theorem 5.1.7 If M
R
for every w , then
R
.
Observation 5.1.1
The classical rules of Modus Ponens and Necessitation fail to be sound with respect to the
rough truth semantics. The rule ,
R
is not sound either, but R2 is sound
and suces for our purpose.
Interestingly, the soundness result establishes that the converse of the deduction theorem
is not true e.g.
R
p Lp, but p ,
R
Lp, p being any propositional variable.
However, the converse does go through, i.e.
R
implies
R
, if
is a modal w, i.e. is of the form L or M, for some w this is because of
DR2.

R
(M)L. So there is no dierence between the modal and non-modal ws in
terms of the object-level implication . However, as just noted, ,
R
L in general
indicating that the meta-level implication
R
does make this distinction.
5.2 Rough belief change
In classical belief revision, the base language is assumed ([12]) to be closed under the
Boolean operators of negation, conjunction, disjunction, and implication. The underly-
ing consequence relation is supraclassical (includes classical consequence) and satises cut,
deduction theorem, monotonicity and compactness.
In contrast, we consider the modal language of L
R
as the base. We follow the classical
line for the rest of the denitions. A belief set is a set K of ws such that Cn(K) = K.
For a pair (K, A), there is a unique belief set K A (K A) representing rough revision
(contraction) of K with respect to A. The new belief set can be dened through a new set
of eight basic postulates (that follow). The expansion K +A of K by the w A is the belief
set Cn(KA). It is expected that rough contraction/revision by two roughly equal [14]
beliefs would lead to identical belief sets. To express this, we make use of the rough equality
connective [11] in S
5
: A B (LA LB) (MA MB).
Postulates for rough revision
(K

1) For any w A and belief set K, K A is a belief set.


(K

2) A K A.
(K

3) K A K +A.
(K

4) If A , K, then K +A K A.
(K

5) If K A is not roughly consistent, then


R
A.
(K

) If K A is roughly inconsistent, then


R
A.
(K

6) If
R
A B, then K A = K B.
(K

7) K (MA MB) K A+B.


(K

8) If B , K A then K A+B K (MA MB).


40 CHAPTER 5. ROUGH BELIEF REVISION
Postulates for rough contraction
(K

1) For any w A and belief set K, K A is a belief set.


(K

2) K A K.
(K

3) If A , K, then K A = K.
(K

4) If A K A, then
R
A.
(K

) If LA K A, then
R
A.
(K

5) K (K A) +A, if A is of the form LB or MB for some w B.


(K

6) If
R
A B, then K A = K B.
(K

7) K A K B K (MA MB).
(K

8) If A , K (MA MB) then K (MA MB) K A.


The major consideration here is to preserve rough consistency during belief change. The
idea, expectedly, is that if K + A is roughly consistent, it could itself serve as K A. Let
us notice the dierence with the classical scenario: suppose K Cn(p), p being any
propositional variable. Then K +p is roughly consistent, and so it is K p itself. But,
classically, K

p K +p. Since we also have the notion of rough inconsistency, there is


the option of avoiding such inconsistency during belief change. It is thus that there are two
versions of postulates involving consistency preservation.
K

1 and K

1 express the constraint of deductive closure. K

2, 3 and K

2, 3 are self-
explanatory. A , K implies consistency and hence rough consistency of K+A, so that, in
view of the previous remarks, K

4 is justiable in the rough context. In K

5, we stipulate
that K A is generally roughly consistent, except in the case when A is roughly valid, i.e.
in no situation denitely A holds (though possibly A may hold). K

4 again stipulates
that, in general, A , K A, except when A is possible in all situations. K

could
appear more relevant: denitely A may follow from our beliefs despite contraction by A
only if A is, in every situation, possible. The controversial recovery postulate K

5 in [6] is
admitted here, only in the case of contraction with a denable/describable [14],[15] belief,
i.e. A such that
S5
A LA. The last two axioms express the relationship of change
functions with respect to beliefs and their conjunctions. The failure of soundness of the
classical conjunction rule A, B
R
A B here, necessitates a modication in the AGM
K

7, K

8, K

7, K

8.
Observation 5.2.1 K

implies K

5 and K

4 implies K

.
The following interrelationships between rough contraction and revision are then ob-
served, if the Levi and Harper identities [12] are used.
Theorem 5.2.1 Let the Levi identity give , i.e. K A (K A) +A,
where the contraction function satises only K

1 8 (K

is not assumed). Then


satises K

1 8, thus K

is also satised.
Proof
(K

1 4) Follow easily.
(K

5) Suppose KA is not roughly consistent. By Theorem 5.1.6, KA


R
B, for any w
B in L
R
. In particular, K A( Cn((KA) A)
R
MB, and K A
R
MB,
for any w B. By using deduction theorem and DR4, K A
R
A. Hence by
K

4,
R
A.
5.2. ROUGH BELIEF CHANGE 41
(K

6) This can be proved by observing that


R
A B if and only if
R
A B, and
by using K

6.
(K

7) First note that K A = K MA, by Observation 5.1.1(c) and K

6 ( is a
special case of ). Thus
K A = K ((MA MB) (MA MB)). *
Secondly, because of DR5, 6 and R2,
K +MA MB = Cn(K A, B), for any K. **
Let K MA MB, i.e. K (MA MB) MA MB
R
. By deduction
theorem, K(MAMB)
R
MAMB . So MAMB K(MA
MB) K K (MA MB) MA MB, the last by K

5, as MA MB
is equivalent to a modal w. It is then easy to see that K (MA MB)
R
MA MB . So MA MB K (MA MB) K (MA MB)
K ((MA MB) (MA MB)), by K

7. Therefore, using *, K A
R
(MA MB) . As (MA MB) is also equivalent to a modal w, by converse of
deduction theorem (cf. Observation 5.1.1(b)), KAMAMB
R
. By (**),
K A A, B(= K A+B)
R
.
(K

8) Suppose B , K A (K A) +A, i.e. K AA ,


R
B. Thus K A
A ,
R
MA MB. Let K A+B = KA+A, B. Using * and K

8, we
have KA = K((MAMB) (MA MB)) = K(MA MB) (MA
MB)) K((MAMB)). Thus, by **, Cn(K((MAMB)) A, B) =
K ((MA MB)) +MA MB = K (MA MB).
Theorem 5.2.2 Let be given by the Harper identity, i.e. K A K K A.
(a) If the revision function satises K

14, K

(K

5 will also be satised) and K

68,
then satises K

1 8 (so K

is satised).
(b) If the revision function satises K

18, then satises K

13, K

and K

58.
Proof
(K

1 3, 6) Follow easily.
((a), K

4) As A K A (by K

2) as well as A K A (assumption), K A is
inconsistent, and hence roughly so. Thus by K

,
R
A. It follows that
R
A.
((b), K

) LA, A K A, by assumption and K

2. So, using S
5
properties, MK
A
S5
LA as well as MK A
S5
LA, implying that K A is not roughly
consistent. By K

5,
R
A. Thus
R
A.
(K

5) Let A be LB, for some w B. If K, it can be shown that A KKA.


By DR2, (K A) +A.
(K

7) Using K

6 we have
K A = K MA = K ((MA MB) MA). *
By * and K

7, KA = K((MAMB)MA) K(MAMB)+MA. Now


let KAKB. Using deduction theorem, K(MAMB)
R
MA .
Similarly, K (MA MB)
R
MB . Observe that
S5
MM M.
Thus by DR7, 8, K (MA MB)
R
(MA MB) M. As K

2 holds and
(MA MB) is equivalent to a modal w, K (MA MB)
R
M, by DR2.
Finally, by DR5, we have K (MA MB).
42 CHAPTER 5. ROUGH BELIEF REVISION
(K

8) Suppose A , K (MA MB) K K (MA MB). So A , K, or A ,


K (MA MB). In the former case, as K

2, 3 already hold, we trivially obtain


K (MA MB) K = K A. Consider the latter case.
Using DR5, (MA) , K (MA MB). By K

8, K (MA MB) + MA
K ((MA MB) MA) = K MA = K A. So K (MA MB)
K (MA MB) +MA K A, and K

8 holds.
5.3 Partial Meet Contraction
In previous section we directly used Levi and Harper identity to investigate rough belief
change. Investigation led us to incorporate some changes in AGM postulates. Since our
results dont lead to a clear picture of the belief change, we start from a basic approach of
constructing contraction function. We can then identify new properties of the constructed
function and hence modify the AGM postulates in a systematic manner.
Similar to classical construction of contraction from maximal subsets of K which fail to
give A, we can construct partial meet contractions in rough background. If the denition
of maximal subsets of K is taken as in section 3.2 then we can show that following holds
Lemma 5.3.1 Let m is in KA for some A in K, then m is closed and Cn(m) = K
where K m.
Proof Let m
R
and assume / m. Now m
R
implies m
R
M which im-
plies M K as m K and K is closed. Hence m m M K, which implies
m M
R
A or m
R
M A. But then m
R
A, as m
R
M and M is modal,
which gives contradiction. Hence m.
For this we already have Cn(m ) K and only K Cn(m ) has to be proved.
Let K m, we show that MA M m and hence any set containing MA and m
(in particular Cn(m)) will imply M(). By contradiction suppose MA M / m
which means m m MA M K as
S5
M (MA M) and (M) K.
Thus (MA M) MA m. But
S5
((MA M) MA) MA gives that
MA(A) m, a contradiction.
Using second part of the above result, we again get that all the beliefs in K outside m
behave identically with respect to m. Suppose m is in KA for some A, and B K m
then addition of some in K m to m will give all of K and in particular B. Thus one can
check that m will satisfy all the three conditions for membership of KB. Thus we have-
Lemma 5.3.2 Let A K and m KA then B K m, m KB.
Using this result we compare K(MAMB), KA and KB. Note that MAMB
R
R
A, B. One can easily see that if m K(MA MB) then it is in KA or KB
because the second condition MA MB / m implies A / m or B / m. What if m is in
KA? The rst and second condition for membership of KMA MB are satised and
our previous lemma implies the third condition - Suppose m m

K, then in addition
to A Cn(m

), Cn(m

) = K is also there implying that B Cn(m

) (as B K). Hence


MA MB is in Cn(m

). We present this result and then move on to new postulates.


Lemma 5.3.3 Let A and B are in K then K(MA MB) = KA KB.
5.3. PARTIAL MEET CONTRACTION 43
In the given background of rough logic we make some changes in the earlier postulates.
The procedure is to use the same denition of contraction using the selection function S as
in equation 3.4 and then identifying the new properties. We consider the classical form of
the recovery postulate, and also nd some changes in K

6, 7, 8
(K

1) K A is a belief set
(K

2) K A K
(K

3) If A / K then K A = K
(K

4) If ,
R
A then A / K A
(K

5) K (K A) +A
(K

6) If A
R R
B then K A = K B
(K

7) K A K B K (MA MB)
(K

8) If A / K (MA MB) then K (MA MB) K A


Theorem 5.3.1 Let be a reexive and transitive relation on M(K). Then function
dened through equation 3.3 with consequence relation
R
, satises (K

1)-(K

8).
Proof
(K

1) This holds because each maximal subset m in S(KA) is closed and intersection
of closed sets is closed.
(K

2) Follows from the denition.


(K

3) If A / K then KA = K and hence K A = K.


(K

4) If ,
R
A then KA is non-empty. Hence A / K A.
(K

5) Let K. Assume ,
R
A so that rst denition of 3.3 is used, we show that
(MA ) K A. Using second part of the lemma 5.3.1 with = A we get
m KA, Cn(m A(orMA)) = K. Thus MA m, K implying
MA K A =

S(KA). Hence (K A) +A.


(K

6) Use, A
R
B and B
R
A implies KA = KB.
(K

7) If
R
A or
R
B then obviously this holds, if
R
A then K A = K and
K B = K (MA MB)(MA MB
R R
B). So let us assume ,
R
A and
,
R
B. We have to show (

S(KA)) (

S(KB))

S(K(MA MB)) . But


S(KAKB) = S(KA) S(KB) S(KA) S(KB). Hence

(S(KA
KB))

(S(KA)S(KB)). Now by property of sets

(S(KA)S(KB)) =
(

S(KA))(

S(KB)) and by lemma 5.3.3 we get (

S(KA))(

S(KB)) =

(S(KA) S(KB))

(S(KA KB)) =

S(KMA MB).
(K

8) Let A / K (MA MB). Thus there exists an m

K(MA MB) such that


m

m, m K(MAMB) and A / m

. Note that m

KA as A / m

. Now we
show

S(K(MAMB))

S(KA) by showing S(KA) S(K(MAMB)).


If m

S(KA) then m

m, m KMA MB implying that m


m, m K(MA MB) due to transitivity of . Hence m

S(K(MA MB))
giving S(KA) S(K(MA MB)).
44 CHAPTER 5. ROUGH BELIEF REVISION
5.4 Rough Consistency
In rough logic we look at following types of consistencies. A subset of L
R
is:
Roughly consistent if M = MA : A is S5 consistent that is there is no B L
R
such that M
S5
B and M
S5
B.
Absolutely consistent if there is a B L
R
such that ,
R
B.
Negation consistent if there is no B in L
R
such that
R
B and
R
B.
Modal negation consistent if there is no B L
R
such that
R
MB and
R
MB.
Note that in S5, consistency is same as classical consistency where absolute and nega-
tion consistency coincide. Also in rough logic negation consistency implies modal negation
consistency but not otherwise.
Lemma 5.4.1 In L
R
, rough consistency, absolute consistency and modal negation consis-
tency are equivalent.
Proof Suppose is not roughly consistent then M is S5 inconsistent, that is to say for
all B L
R
M
S5
B. Hence M
S5
MB and we have
R
B for all B in L
R
that is
to say is not absolutely consistent. If is not absolutely consistent then obviously it is
not modal negation consistent. Finally suppose is not modal negation consistent. Thus

R
MB and
R
MB for some B L
R
. Hence M
S5
MMB and M
S5
MMB,
which is same as M
S5
MB, MB. Thus M is not S5 consistent or is not roughly
consistent.
Lemma 5.4.2 Let L
R
.
(i) A
R
B, B L
R
if and only if
R
MA.
(ii) MA
R
B, B L
R
if and only if
R
A.
Proof (i) if part follows directly from equivalence of absolute consistency and modal nega-
tion consistency. For only if part A
R
B if and only if M MA
S5
MB.
Thus in particular M MA
S5
MA. Using D.T., M
S5
MA MA, but
MA MA
S5
MA and hence M
S5
MA(MMA). Hence
R
MA.
(ii) Replacing A by MA in rst part we get MA
R
B, B L
R
if and only if

R
MMA. Since MMA LMA MA and MA
R R
A, the result follows.
5.5 Maximal Consistent Sets
In rough logic maximal consistent sets are those roughly consistent sets which become not
roughly consistent on adding anything outside them. Formally, M
r
L
R
is maximal con-
sistent if:
(i) M
r
is roughly consistent and
(ii) If A / M
r
then M
r
A is not roughly consistent.
Following properties of rough maximal consistent sets are notable:
(a) M
r
is closed : Suppose M
r

R
A andA / M
r
. Thus M
r
A is not roughly consistent.
Hence MM
r
MA is S5 inconsistent, that is MM
r

S5
MA. But then M
R
A
gives MM
S5
MA which with MM
r

S5
MA implies that M
r
is not roughly
consistent - a contradiction.
5.6. REVISION FROM SYSTEM OF SPHERES 45
(b) For every B L
R
exactly one of MB and MB is in M
r
: Obviously both MB
and MB can not be in M
r
as it is (also) modally negation consistent. Now suppose,
MB / M
r
. Thus MM
r
MMB(or MB) is not S5 consistent and we have MM
r

S5
MB(or MMB). Which implies M
r

R
MB or MB M
r
as M
r
is closed.
(c) For every B L
R
exactly one of LB and LB is in M
r
: Taking B as B in (b), we
get exactly one of LB(MB) or LB(MB) is in M
r
.
(d) Every roughly consistent set has a maximal consistent extension : Let
0
,
1
,
2
,
3
, ...
be an enumeration of L
R
. Construct
0
,
1
, ...,
i
, ... as:

0
:=

i+1
:=
i

i
if
i

i
is roughly consistent,
i
otherwise. Take M
r
=

i0

i
.
Note that each
i
is roughly consistent. Since M
r
, we only need to prove
that M
r
is maximal consistent. First M
r
is roughly consistent: Suppose not then
MM
r

S5
B, B, thus by compactness there is a nite subset of M
r
which gives both
B and B. Now if k is the maximum index of formulae contained in the nite subset
then M
k+1

S5
B, B meaning
k+1
is not roughly consistent - a contradiction. For
proving second condition, let A / M
r
. So A =
i
for some i and
i+1
does not contain

i
. Thus
i+1
=
i
and
i

i
is not roughly consistent. Thus M
r

i
is also
not roughly consistent as
i
M
r
.
(e) If K is closed and roughly consistent then K =

[K] : Since K is roughly consistent
[K] is non-empty, thus obviously K

[K]. For the other side, we show that for all


A / K, A /

[K]. We rst show K+MA is roughly consistent : If K+MA is not


roughly consistent then MK MMA(or MA) is S5 inconsistent implying that
MK
S5
MA which means K
R
A or A K - a contradiction. Now maximal
consistent extension of K +MA is in [K] and does not contain MA(or A) implying
that A /

[K].
(f) [MA MB] = [A] [B] : Straight from the fact that MA MB
R R
A, B.
5.6 Revision from System of Spheres
Using system of spheres we can obtain revision function. We dene the revision as:
K A =

C(A), if [A] ,=
K

where, C(A) = S
A
[A] (5.1)
Note that [A] is empty means A has no maximal consistent extension. That is possible only
when A is not roughly consistent which is same as
R
MA. Based on new properties of
maximal consistent sets we make changes in the classical postulates for revision. Following
postulates t best in rough context:
(K

1) K A is a belief set
(K

2) A K A
(K

3) K A K +A
(K

4) If MA / K then K +A K A
(K

5) K A = K

if and only if
R
MA
(K

6) If A
R R
B then K A = K B
(K

7) K (MA MB) (K A) +B
(K

8) If MB / K A then (K A) +B K (MA MB)


46 CHAPTER 5. ROUGH BELIEF REVISION
We prove that revision function dened in equation 5.1 satises all the above postulates.
Theorem 5.6.1 Let K be a belief set and its revision with respect to a belief A is dened
as in equation 5.1 then it satises all the above postulates.
Proof
(K

1) C(A) contains maximal consistent sets which are closed. So there intersection has
to be closed, and hence K A is a belief set.
(K

2) If [A] is non-empty then each maximal consistent set in C(A) is also in [A] and hence
contains A, otherwise when [A] is empty K A = K

which contains A.
(K

3) If K +A = K

then obviously K A K +A, otherwise when K +A ,= K

then
it is absolutely consistent and hence roughly consistent. Thus K + A has a maximal
consistent extension M
r
. Since M
r
contains both K and A, [K] [A] ,= and we
have S
A
= [K]. Thus C(A) = [K] [A]. But then [K +A] [K] [A] implying that
K A =

C(A) =

([K] [A])

[K +A] = K +A.
(K 4) If MA / K then K + A is roughly consistent which has a maximal consistent
extension. Thus S
A
= [K] and we have C(A) = [K] [A]. With [K] [A] [K +A],
we get K +A =

[K +A]

([K] [A]) =

C(A) = K A.
(K 5) if By denition K A can be K

only when
R
MA.
only if Using K

2 we have A(or MA) K A. Thus with MA K A, we


have K A is not modally negation consistent which implies K A is not absolutely
consistent. Thus K A = K

.
(K

6) Use A
R R
B implies [A] = [B].
(K

7) Let K (MA MB). We show that MB M M


r
for all M
r
C(A), so
that M(or )

C(A) +B = K A+B. Suppose M


r
C(A), if MB M
r
then
with
S5
MB (MB M) we have MB M M
r
. If MB / M
r
then
MB(or B) M
r
and hence M
r
[A] [B] = [MAMB]. But M
r
C(A) = S
A
[A]
also implies that M
r
S
A
. Thus S
A
[MA MB] ,= and we have S
MAMB
S
A
,
but MA MB
R
A implies that S
A
S
MAMB
. Thus we get S
MAMB
[MA
MB] = S
A
[MAMB] which implies M
r
S
MAMB
[MAMB] = C(MAMB).
Hence M
r
and again we get MB M M
r
.
(K

8) Let K A + B. Thus MB K A. Now MB /



C(A) implies
there exists an M
r
S
A
[A] such that MB / M
r
. Hence B M
r
, using a
similar argument as in proof of K

7, we get S
A
[A] S
MAMB
[MAMB]. But
MB KA implies MB M
r
for all M
r
S
A
[A], hence MB M
r
for all M
r
S
MAMB
[MA MB] = C(MA MB). Since B M
r
for all M
r
in
[MA MB] we get that M
r
for all M
r
in C(MA MB).
5.7 Epistemic Entrenchment
In rough context because of A, B ,
R R
A B the epistemic entrenchment ranking
denition must be changed. EE2 says that if A
R
B then B A, thus we have A and
B MA MB A B. Now EE3 can be left as it is or we can replace A B by
5.7. EPISTEMIC ENTRENCHMENT 47
MA MB. But then rst option makes MA MB =
E
A B, where C =
E
D means
C D and C D, and hence we choose the second option. Other postulates seem to be
fundamental to epistemic preference and left unchanged.
(EE1) is transitive.
(EE2) If A
R
B then B A
(EE3) Either MA MB A or MA MB B
(EE4) K ,= K

then B A for all B if and only if A / K


(EE5) If A B for all B then
R
A.
Note that :
(i) Since MA
R R
A, MA =
E
A.
(ii) A B =
E
M(A B) =
E
MA MB.
(iii) A
R
B implies A C
R
B C.
(iv) (MA MB) (MA MC)
R R
MA (MB MC).
Given an epistemic entrenchment ranking , contraction function can be dened as
K A =

B K : A < A B, if ,
R
A
K, otherwise
(5.2)
A similar method can be used in the reverse way. Given a contraction function an epistemic
entrenchment ranking can be obtained as
A B if A / K (A B) or
R
A B (5.3)
Theorem 5.7.1 Let be a contraction function and let K is a belief set. Then the ordering
dened in 5.3 is an epistemic entrenchment ranking.
Proof Follows on the classical lines.
Theorem 5.7.2 Let K be a belief set. For every contraction function of K there is an epis-
temic entrenchment (related to K) such that 5.2 is true. Conversely for every epistemic
entrenchment related to K, there is a contraction function satisfying 5.2.
Proof For the rst part theorem 5.7.1 gives the required epistemic entrenchment ranking.
We only need to show that 5.2 holds for this ranking.
If ,
R
A then B K : A < A B = B K : A / K A (A B) = K A and
(A B) K A. We will show that if B K then (A / K A and (A B) K A)
is same as (A / K A and B K A) which is same as denition of K A. If B K
then using K

5 we have MB (K A) +A which implies MA MB K A. Thus if


A B K A then (MA MB) (MA MB) = MB(or B) K A. If B K A
then obviously (A B) K A.
If
R
A then A K A. That is (K A) + A = K A. Using (K

5) we have
K (K A) +A = (K A). Along with (K

2) we get K A = K.
For the second part we can dene the function as in equation 5.2. We only need to show
that this function is a contraction.
48 CHAPTER 5. ROUGH BELIEF REVISION
(K

1) Suppose K A
R
for some L
R
. By compactness of
R
there exist

1
,
2
, ..,
n
such that
1
,
2
...
n

R
that is to say M
1
M
2
... M
n

R
.
To show K A we have to show that K and either A < A or

R
A. Suppose ,
R
A, then A
i
=
E
MA M
i
> A for all
i
. Hence we have
MA(M
1
M
2
.. M
n
) > A as by (EE2) MA(M
1
M
2
.. M
n
) =
E
(MA M
1
) .. (MA M
n
) which is equal to some MA M
k
(because
of (EE3)) and for which MA M
k
=
E
A
k
> A. Now (EE2) also gives
MA (M
1
M
2
.. M
n
) MA M =
E
A which with (EE1) gives
A < A . So for n 1 above proof works. If n = 0 then
R
hence
R
A and
by (EE2), C A for all C L. ,
R
A and (EE5) give C > A for some C. (EE1)
now gives A > A.
(K

2) Follows from the denition.


(K

3) As A / K, K ,= K

. Hence (EE4) gives for all B L, B A. Now if K then


(EE4) gives a C such that , C or > C A. Using (EE1) and (EE2) we have
A > A or A > A.
(K

4) A
R
A hence A =
E
A which means that A A > A is not possible.
(K

5) Let K, we show that MA M KA. By the equation 5.2 we need to prove


that MA M K and either A < A (MA M) or
R
A. MA M K
as K. Let us assume that ,
R
A and show A < A (MA M). Since

R
MA(MA M), by (EE2) we have C MA(MA M) for all C L
R
.
A(or MA) not being a theorem satises A < C for some C by (EE5). Thus by (EE1)
we get A < MA (MA M) =
E
A (MA MB).
(K

6) If A
R R
B then by (EE2) we have A =
E
B and A C =
E
B C. Thus
A < A C is same as B < B C. The equality of K A and K B now follows
from the denition.
(K

7) We have to show that KAKB K(MAMB). Let KA and KB


that is A < A and B < B. Thus by (EE2), (EE1) we get MAMB < A
and MA MB < B . Now (EE3) gives that M(A ) M(B ) =
E
(A )
or (B ). Using (EE1), AB < M(A) M(B ) =
E
(MAMB) M i.e.
M() KMAMB. The last equality followed from the fact : M(AB)
R R

MA MB.
(K

8) A / K(MAMB) implies MAMB (MAMB)A =


E
(MAMB)MA =
E
A. Now if K(MAMB) then MAMB < (MAMB) . Thus by (EE1)
we get (MA MB) > A. Finally using (EE2) we get A =
E
(MA M)
(MA M) (MB M) =
E
(MA MB) M =
E
(MA MB) > A, hence
A > A and we have K A.
5.8 Interrelations
We have changed contraction and revision postulates independently, based upon partial
meet and spheres. So here based upon the changed postulates, we establish new relations
between revision and contraction. Revision of a belief set K with respect to a sentence
A can be seen as removal of any incompatible information and then addition of A along
5.8. INTERRELATIONS 49
with all logical consequences. We nd that among MA, A, LA only LA MA
is incompatible with A that is to say MA, A is not roughly consistent. In our earlier
approach where we removed A, MA was automatically removed because MA
R
A.
But removal of A is not in line with our principle of minimal change, as A is safe with
respect to A. Our new Levi-identity is -
K A = (K MA) +A (5.4)
Let us call the revision function above R(). As usual following theorem tells that func-
tion R() dened through Levi - identity is a revision function according to new postulates.
Theorem 5.8.1 If a contraction function satises (K

1) to (K

6) then R() satises


(K

1) to (K

6). Furthermore if also satises (K

7) and (K

8) then R() satises


(K

7) and (K

8) respectively.
Proof
(K

1) K A = (K MA) +A is obviously closed.


(K

2) A K A as (K MA) A
R
A.
(K

3) From (K

2), K MA K or (K MA) A K A. Thus Cn((K


MA) A) Cn(K A). Hence K A K +A.
(K

4) Let MA / K, using (K

3) we have K MA = K. Hence Cn(K A) =


Cn((K MA) A) or K +A = (K MA) +A = K A.
(K

5) if Let
R
MA. Thus MA and A(MA) (KMA)+A. Thus (KMA)+A
is not modal negation consistent. Hence (KMA) +A is not absolutely consistent
and is equal to K

.
only if Let (K MA) + A = K

. Thus from lemma 5.4.2. we get K MA


R
MA. (K

4) now implies
R
MA.
(K

6) Let K A = (K MA) + A. D.T. gives (K MA)


R
(MA ). But
A
R R
B implies MA
R R
MB, hence using (K

6) we have (K MB) =
(K MA)
R
(MA ). Thus (K MB) + B
R
(MA ) which in turn
yields (K MB) + B
R
as B
R
A(MA). So, K A K B. Similarly
K B K A giving K A = K B.
(K

7) Let (M) K MA MB = (K M(MA MB)) + (MA MB). Thus


(MAMB) M (K(MAMB)). Note that MA
R R
(MAMB)
(MAMB). We show that (MAMB) M K (MAMB) as then by
(K

7) and (K

6), K(MAMB) K(MAMB) K((MAMB)


(MAMB)) = KMA. Which will give (MAMB) M KMA. And
then as MA MB
R R
A, B, M() ((K MA) +A) +B = K A+B.
So (MA MB) M K (MA MB) needs to be proved. Using (K

2) and
(K

5) we have (MA MB) M K (MA MB) K K (MA


B) +(MAB). Thus ((MAMB) M) K (MAMB) +(MAMB)
which on application of D.T. gives ((MA MB) M) K (MA MB) as
((MA MB) ((MA MB) M))
R
((MA MB) M).
50 CHAPTER 5. ROUGH BELIEF REVISION
(K

8) Let MB / K A = K MA+A and K A+B = ((K MA) +A) +B.


Thus (MA MB) / KMA and (MAMB) KMA = K(MA
MB) MA. Now using (K

8) with A

= MA MB and B

= MA we get
(MAMB) K(MA MB) MA K(MA MB). Thus nally
K(MA MB) +(MAMB) = K(MAMB) as (MA MB)
R R

(MA MB).
Similar to levi identity, we change the Harper-identity.
K A = K K MA (5.5)
We call the function above C(). Following theorem ensures that C() is indeed a
contraction function.
Theorem 5.8.2 If a revision function * satises (K

1) to (K

6) then C(*) satises


(K

1) to (K

6). Furthermore, if * also satises (K

7) and (K

8) then C(*) satises


(K

7) and (K

8) respectively.
Proof
(K

1) K A = K K MA is a belief set as intersection of two belief sets is also a


belief set.
(K

2) K A = K K MA K.
(K

3) Let A(MA) / K i.e. MMA LMA MA / K. Using (K

4), K MA =
K + (MA). But K K + (MA) giving K A = K K + (MA) = K.
(K

4) Let A K A = K K MA. Thus A K MA. Also, using (K

2),
MA K MA and we have K MA = K

. (K

5) now gives
R
MMA or

R
A.
(K

5) Let K, we need to show (M) (KKMA)+A.


R
M (MA M)
implies MA M K. Thus its sucient to prove MA M K MA. But
that is ne as
R
MA (MA M) and MA K MA because of (K

2).
(K

6) If A
R R
B then MA
R R
MB. From (K

6) we have KMA = KMB


which gives K A = K K MA = K K MB = K B.
(K

7) Let K A K B, thus K, K MA, K MB. Taking A

= (MA
MB) and B

= (MA) and using (K

7) we get K ((MA MB) MA) =


K MA K (MAMB) +MA - note that A

, B

are modal, and hence are


equivalent to MA

, MB

respectively. Similarly we can get K MB K (MA


MB) + MB. Using D.T., MA M, MB M K (MA MB).
Now C() already satises (K

5) so M K (K K (MA MB)) + MA
MB. And we have MA MB M K (MA MB). But MA MB
M, MA M, MB M
R
M and hence K (MA MB). Thus
K K (MA MB) = K (MA MB).
(K

8) Let A(MA) / K(MAMB). Which means MA / K or MA / K(MAMB).


If MA / K then K(MAMB) K = KMA. So consider the case when MA K
but MA / K (MA MB). Taking A

= MA MB, B

= MA and using
5.8. INTERRELATIONS 51
(K

8) we have K (MAMB) +MA K (MAMB) MA = K MA.


But K (MAMB) K (MAMB) +MA and we have K(MAMB) =
K K (MA MB) K K MA = K MA.
Chapter 6
Conclusions
We started with a survey of the current techniques for reasoning with inconsistent and
imprecise knowledge. Starting from the initial survey on belief change functions we moved
on to implementation of these functions. The implementation actually helped us in getting
hold of practicality of belief change. Soon we realized the restriction posed by classical
belief change and its limited application to day-today belief change. We identied two main
deciencies: rst is lack of expression through beliefs and second is restrictiveness with re-
spect to contradiction. So we moved onto exploring the internal structure of belief change,
and looked into classical results and worked out proofs of all the important concepts epis-
temic entrenchment, partial meet contraction and system of spheres in belief change. After
getting conversant with the logical and structural properties of belief change we thought of
incorporating more expressibility and flexibility. We found existing approaches of doing
belief change in non-classical logics very rudimentary as far as practicality of the approaches
is concerned.
We analyzed belief change in 3-valued logic. From the interpretations we nd that more
truth values lead to more variety of beliefs, with each values specifying varying level of
possibility of a belief. Like MA (A A) means possibly A is true. We then extended
our approach to an n-valued logic, generalization of 3-valued belief change to n-valued
belief change then led us to conclude that formulae of the type (A (A...(A A)))
represent possibility of A. As the number of As increases possibility of A reduces. Thus
we very much achieve expressibility goal in n-valued logics.
By exibility we meant exibility with respect to contradiction. In day today life one
can have all sorts of information and in particular both A and A, but as agent might not be
sure of credibility of the dierent sources of the information he is unable to keep just one of
them as his belief. This led us to use rough logic for belief revision. In rough logic anything
in the belief set which is not of the form MA or LA is regarded as a belief whose credibility
is still under suspicion. Thus conjunction of these beliefs with their negation dont lead
to a contradiction. We can look at rough truth as soft truth where initially a belief is
present in its possibility form and gradually with accumulation of more knowledge changes
to its actual truth. We used two independent approaches for constructing contraction and
revision functions. After identifying the properties we made the corresponding changes in
the AGM postulates. We very much achieve the goal of exibility, as any non-modal belief
A is actually only a possibility assertion that can be actually true or false.
52
6.1. SCOPE FOR FUTURE WORK 53
6.1 Scope for Future Work
As we see now, there can be plenty of approaches of doing belief revision in non-classical
logics. But one should try to rst identify the desired properties of the belief change and
then invent a logic where those properties can be achieved. Like one can think of rough logic
where there are more than one possibility notions of a belief such that with accumulation of
more information belief moves up the possibility hierarchy. Probably a logic which mixes up
the n-valued and rough logic appropriately is a good choice for this kind of belief revision.
We have implemented 3-valued belief revision using tertiary decision diagrams. So one
can also think of implementation in rough logic. Foremost concern in such an implemen-
tation would be designing an automated theorem prover. Other issues originating from
implementation can be really interesting for example, free belief revision, where initially
agent does not specify any preference, and the implementation itself produces all the pos-
sibilities. The agent then chooses the one which ts most to his current state. Gradually
these kind of interactions makes the implementation trained enough so that at times it can
make choices by itself. Currently I am looking at these ideas.
Bibliography
[1] Grigoris Antoniou. Nonmonotonic Reasoning, chapter 14,15. The MIT Press, Cam-
bridge, Massachusetts, 1997.
[2] M. Banerjee. Rough sets and 3-valued lukasiewicz logic. Fundam. Inf., 31(3-4):213220,
1997.
[3] M. Banerjee. Rough truth, consequence, consistency and belief revision. In LNAI 3066:
Proc. 4th Int. Conf. On Rough Sets and Current Trends in Computing (RSCTC2004),
pages 95102, Uppsala, Sweden, June 2004, 2004. Springer-Verlag.
[4] D. Becchio. Logique trivalente de lukasiewicz. In Ann. Sci. Univ. Clermont-Ferrand,
volume 16.
[5] V. Boicescu, A. Filipoiu, G. Georgescu, and S. Rudeano. The wajsberg axiomatization
of the 3-valued logic, Lukasiewicz-Moisil Algebras, chapter 9. 1991.
[6] Alchourron C.A., Gardenfors P., and Makinson D. On the logic of theory change:
partial meet contraction and revision functions. The Journal of Symbolic Logic, 50:510
530, 1985.
[7] Restall G. and Slaney J. Realistic belief revision. In Proc. 1st World Congress in the
Fundamentals of Articial Intelligence, pages 367378, Paris, July 1995.
[8] Peter Gardenfors. Knowledge in Flux:Modeling the Dynamics of Epistemic States,
chapter 3,4. MIT Press, Bradford Books, Cambridge, MA, 1988.
[9] Adam Grove. Two modellings for theory change. Journal of Philosophical Logic,
17:157170, 1988.
[10] Goubault-Larrecq Jean and Mackie Ian. Proof Theory and Automated Deduction, vol-
ume 6 of Applied Logic Series. Kluwer Academic Publishers, may 1997.
[11] Banerjee M. and Chakraborty M.K. Rough consequence and rough algebra. In Rough
Sets, Fuzzy Sets and Knowledge Discovery, Proc. Int. Workshop on Rough Sets and
Knowledge Discovery (RSKD93), pages 196207. Springer-Verlag.
[12] G ardenfors P. and Rott H. Belief revision. In Handbook of Logic in AI and Logic
Programming: Epistemic and Temporal Reasoning, volume 4, pages 35132. Clarendon,
1995.
[13] Mary-Anne Williams. Changing nonmonotonic inference relations. In WOCFAI, pages
469480, 1995.
54
BIBLIOGRAPHY 55
[14] Pawlak Z. Rough sets. Int. J. Comp. Inf. Sci., 11:341356, 1982.
[15] Pawlak Z. Rough logic. Bull. Polish Acad. Sc. (Tech. Sc.), 35(5-6):253258, 1987.

Vous aimerez peut-être aussi