Transcript B - SSDI

Belief Revision
Lecture 2: Beyond AGM
Gregory Wheeler
[email protected]
Outline
•
•
•
•
•
Withdrawal vs. Contraction
Belief Bases
Updates
Choice
Uncertainty
Belief Bases
In AGM, a belief state is modeled as a
deductively closed theory. Thus, there is no
distinction between ‘basic beliefs’ and
‘derived beliefs’.
Thus, retracting a formula  does not entail
retracting all formulas derived from .
Belief Bases
Example:
Let B = {(Match  Campfire), Match}.
Let K = Cn(B).
Notice: Campfire  (K - Match)
This is an example of the coherence approach to theory
change, where the maximum amount of information is
preserved.
Belief Bases
A belief base is a finite set of nondeductively closed formulas.
B1 = {,   }
B2 = {, }
Note that B1  B2, but Cn(B1) = Cn(B2).
Belief Bases
A Foundationalist approach to theory
change maintains that changes are only
made to belief bases, not to deductively
closed theories.
B1 = {,   }
  (B1 -- )
B2 = {, }
  (B2 -- )
Note: Base revisions and contractions have properties
different than AGM style operators.
Updates vs. Revision
The difference lies in the source of incorrect beliefs
Belief revision assumes that the world is unchanging: an
agent’s change in belief occurs when that agent discovers
something new about the static world or discovers an error
in his beliefs about the static world.
Belief updates involve a change in belief prompted by
actions in a dynamic world. This process behaves
differently than revision.
Updates vs. Revision
Suppose:
I believe that Reema lives in Rochester.
Types of changes:
1. I learn that she has just moved to Cairo.
(Update)
2. I learn that I have been mistaken and that she lives in
Cairo.
(Revision)
Changes in the World: The idea
• Interpretation of a belief set B
– the set of possible worlds where B is true
• notification of some change in the actual world
• The agent’s description of the possible states of
affairs must be modified accordingly:
– Our description of the actual world is typically incomplete, which
means that there are several states of affairs (possible worlds) that
are consistent with what we believe. Hence, an update must ensure
that the changes are made true in the “candidate worlds” that
survive the update.
Selection Functions
• Interpretation of a belief base B:
B = {w : w is a possible world, and w  B}
= [B]
Selection Functions
• Interpretation of a belief base B:
B = {w : w is a possible world, and w  B}
= [B]
• Update of a world w  [B] by 
= the possible worlds w close to where  is true
[]
= w  [] (selection function)
w
Selection Function Semantics
Update B by  = [B  ] = w  [B]w  []
[B]
[B  ]
w  []
w
v
v  []
Updates vs. Conditionals
• Update Problem:
– current belief base B
– input 
– a belief 
• Ternary Relation R(B, , ); algebraically:
1. R(B, , ) iff B     (B updated by  entails )
B   “B has been updated by ” used in databases
2. R(B, , ) iff B 

 (B entails  if updated by )
   “if one updated with  then  follows” (counterfactuals)
The Ramsey Test
B   –  iff B – 


If two people are arguing ‘If p will q’ and are both in doubt as to p,
they are adding p hypothetically to their stock of knowledge and
arguing on that basis about q.
-Frank Ramsey 1931
This is how to evaluate a conditional: First, add the antecedent
(hypothetically) to your stock of beliefs; second, make whatever
adjustments are required to maintain consistency (without modifying
the hypothetical belief in the antecedent); finally, consider whether or
not the consequent is true.
-Robert Stalnaker 1968
Gärdenfors Impossibility Theorem
Theorem (Gärdenfors 1978): Suppose
•  is a metalanguage belief change operator, mapping
bases and target formula to new bases;
• L is a logic for an object language 
•  and  satisfy the Ramsey test:
B   –  iff –L B  


• the logic is ‘rich enough’;
Then,  does not satisfy the AGM postulates.
Gärdenfors Impossibility Theorem
Remarks:
• The result does not depend upon –, and the
theorem holds even if – is substructural or
non-monotonic (Makinson 1990).
• Similar to Gödel’s incompleteness result, no
object language predicate can express consistency.
• Result does not hold for updates (Grahne, KR
1991).
Language vs. Metalanguage
• Traditionally
–  metalanguage operation
• single model
• non-constructive postulates
“If o is consistent, then y -d o is consistent”,…
–

object language operator
• class of models
• axiomatics (constructive
Language vs. Metalanguage
• Now:
– two dyadic modal operators in the object
language:
• update operator: 
• conditional operator: 
Logics based on the Ramsey Test
• classical logic, plus
• two inference rules, “the Ramsey Rules”
B    
B    
B    
B    
where ‘B’ is the conjunction of the elements (wffs) in the
belief base.
Updates and conditionals as inverse modal
operators
(Ryan and Schobbens, JoLLI 1997)
Switching Lemma (de Rijke, Venema, Roorda, Dunn)
• The conversion axioms
B    (B  )
(  )   
can be derived from the Ramsey rules
• …and vice versa
(cf. tense logics, dynamic logic with inverse modality,
program logics.)
Derivable principles for updates -1
•
B1  B2
B1    B2  
(syntax-independence)
•
B1  B2
B1    B2  
(monotony)
• (B1    B2   )  (B1  B2) 
(updating disjunctions)
•     
(inconsistency preservation)
Derivable principles for updates -2
•
Example: proof of     
1.
2.
    
    
by classical logic
by the Ramsey Rules
Derivable principles for updates -3
•
Example: proof of
B1  B2
.
B1    B2  
1.
2.
3.
4.
5.
B1  B2
B1   
B2    
B1    
B2    
hypothesis
from 1, with new 
B2    
B1    
from 3
by a proof similar to 1-4.
from 2, by the Ramsey rules
Derivable principles for conditionals
(Wansing, JLC 1994)
.
2.
1  2
1.

 1  

2


3.


(1 2)  (
4.


T

1  

1  2
 1  


.
2
2)
1-4 are axioms of semi-normal conditional logic Ck (Chellas, JPL 1975)
Recovering normal conditional logics
CK
= (Conditional, Kripke)
1  2
= Ck +
1 []   2 [] 
Selection function semantics:
M = W, ·, V such that:
• W

• · :Wx2 2
• V(w)  ATM for all w  W.
(set of possible worlds)
(selection function)
(truth assignment)
[  ] = {w  W : w · []  []}
Thm (Chellas, 1975): sound, complete, finite model prop., decidable.
Conversely
• Given:
– a conditional logic L
– its tense extension L, by the Ramsey rules
• then we may ask which properties does L inherit?
(completeness, decidability, finite model, correspondence)
• ALL if L is one of the standard systems
• Correspondence only in the general case
(Wolter, NDJFL 1998)
Normal Update Logics
UCK = (Update Conditional, Kripke)
= CK + the Ramsey Rules
1  2
1 []   2[] 
1  2
= classical logic + Ramsey rules +
  1    2
= classical logic
+ Ramsey rules +

Thm (Herzg 1998): sound and complete for selection
function semantics

Four desiderata for belief update
• syntax independence
if 1 2 then B  1  B  2
if B1 B2 then B1    B2  
• success
B   – 
• “maximum” consistency
B   is consistent iff both B and  are consistent.
• minimal change
B   is “as close to B as possible”
(Note: the last 3 require extensions of UCK.)
UCK and update desiderata
• success
– axiomatically: UCK + (B  )  
– semantically: w · []  []
• consistency
B    
UCK
+
– axiomatically:
B 
– semantically: w · []  , for all w  W, for all
consistent formula  of classical logic.

UCK and update desiderata
• minimum change
– axiomatically: difficult to express if B   is inconsistent
– semantically: minimal change “defined away” by selection
function
• weaker forms of minimal change
B 
UCK +
(idempotence)
B    B
minimality ~ orderings

Cumulative update logics
Minimal change: orderings of closeness associated to worlds.
[]
w
w · []
To update by o, go to worlds where o is true and stop--these are the
closest worlds.
Cumulative update logics - semantics
• M = W, {≤w : w  W}, V such that
– W a set of possible worlds
– ≤w partial pre-order (transitive, reflexive relation) on W
u ≤w v = “u is at least as close to w as v is”
• Sw = {u  W : there is v  W such that u ≤w v}
= accessible worlds
• w · [] = the -worlds closest to w
= min≤w([]  Sw)
= {u  []  Sw : v  []  Sw, if v ≤w u then v ≤w u}
• truth conditions as for UCK
if W is infinite: Limit Assumption entails that []   implies w · []  .
Cumulative conditional logics -axiomatics
CL = (Conditional, Lewis)
= CK
+   
+ ((1  )  (2
+ ((1  2)  (2
))  ((12)  )
 1)) 
((1  )  (2  ))

(Lewis 1973; Burgess 1981)
Cumulative conditional logics -axiomatics
UCL
= (Update, Conditional, Lewis)
= UCK + CL
= UCK
+ (B  )  
+ (B  (1  2))  (B  1)  (B  2)
(B  1 )  2,(B  2 )  1

(B  1)  (B  2 )



UCL: derivable principles
• cumulativity
 (1 [] 2 ) ((1 []) ((1  2 )[]))
B  1  2

(B  1)  (B  (1  2 ))
KM theory
(Katsuno and Mendelzon, 1991)
A KM-model is a UCL-model M = W, {≤w : w  W}, V s.t.:
• W = 2ATM the set of all logically possible worlds
= all truth assignments
• ≤w is a ‘faithful’ partial order:
– w ≤w u for all u
– if u ≤w w then u = w.
(hence: min≤w (W) = {w})
• V(w) = w
(every w is a set of atoms)
KM theory
• The formula ‘B  ’ denotes the
replacement belief base resulting from an
update of B by .
• The formula  here denotes a new fact
observed by an agent in response to a
change in the world.
KM Update Postulates
(U1) (B  )  .
Updating a belief base by a target formula yields a
replacement belief base that implies the target
formula.
(U2) If B   then (B  )  B.
If a belief base implies a target formula , then updating by
 yields a replacement belief base that is logically
equivalent to the original belief base. (problematic)
KM Update Postulates
(U3) If B and  are consistent, then (B  ) is
consistent.
Updating a consistent belief base by a consistent target
formula yields a consistent replacement belief base.
(U4) If    , B1  B2, then (B  )  (B 
).
Logically equivalent belief bases updated by logically
equivalent target formulas yield logically equivalent
replacement belief bases.
KM Update Postulates
(U4.1) If B1  B2, then (B1  )  (B2  ).
(U4.1) If   , then (B  )  (B  ).
Postulates of syntax independence: U4.1 and U4.2 together hold that if
logically equivalent belief bases are updated by logically equivalent
target formula yield logically equivalent replacement belief bases.
KM Update Postulates
(U5) ((B  )  )  (B  (  )).
The conjunction  and the updated belief base B by a
target formula  implies the update of B by the conjunction
  .
(U6) If (B  1)  2 and (B  2)  1, then
(B  1)  (B  2)
If an update by , implies  and an update by  implies ,
then an update by  is equivalent to an update by .
KM Update Postulates
(U7) If B is complete, then
((B  1)  (B  2))  (B  (1  2)).
Updating by the target formula 1 and updating by the
target formula 2 entails an update by the target formula
(1  2). This requires a finite language.
(U8) (B1  B2)    (B1  )  (B2  ).
Updating one or another belief base by a target formula is
logically equivalent to updating one belief base by a target
formula or updating the other by the target formula.
KM theory
Remarks:
Success (U1); syntax independence (U4.1) + (U4.2);
consistency ‘as much as possible’ (U3).
Minimal change? Not always satisfied.
KM theory
• a faithful partial pre-ordering for maximal change:
≤w = {w,u : w  W}
• consequence:
min≤w([]) = {w}, if w  []; otherwise [].
• the maximal change operation

B if B | 
B    
 if B does not entail 
satisfies the KM postulates.

References -1
• John Burgess, “Quick completeness proofs for some logics of
conditionals”, Notre Dame Journal of Formal Logic, 22:76-84, 1981
• Brian Chellas, “Basic conditional logics”, Journal of Philosophical
Logic, 4:133-153, 1973.
• Peter Gärdenfors, “Conditionals and changes of belief”, in I.
H\Niiniluoto and R. Tuomela (eds.), The Logic and Epistemology of
Scientific Change, vol. 30, Acta Philosophica Fennica, 1978.
• G. Grahne, “Updates and counterfactuals”, Proc. KR 1991.
• Andreas Herzig, “Logics for belief base updating”, in Handbook of
defeasible reasoning and uncertainty management, vol. 3. D. Gabbay
et. al. (eds.), Kluwer Academic Publishers, 1998.
References -2
• Hirofumi Katsuno and Alberto O. Mendelzon, “On the difference
between updating a knowledge base and revising it”, in Belief
Revision, P. Gärdenfors (ed.), Cambridge University Press, 1992. (First
version appearing in Proc. KR 1991).A.
• A.M Keller and M. Winslett, “On the use of an extended relational
model to handle changing incomplete information”, in IEEE
Transactions on Software Engineering, 1985.
• David Lewis, Counterfactuals. Basil Blackwell, Oxford 1973.
• David Makinson, “The Gärdenfors Impossibility Theorem in nonmonotonic contexts”, Studia Logica, 49:1-6, 1992.
• Frank P. Ramsey, Collected Essays of F. P. Ramsey, Oxford Press.
• Robert Stalknaker, “A theory of conditionals”, Studies in Logical
Theory. Blackwell, Oxford, 1968.
References -3
• H. Wansing, “Sequent Calculi for Normal Modal Propositional Logics”
, Journal of Logic and Computation, 4:125-42, 1994.
• F. Wolter, “A counterexample in tense logic”, Notre Dame J. of Formal
Logic, 37(2), 1997. (Special issue on combining logics).