Upload
others
View
9
Download
0
Embed Size (px)
Citation preview
Department ofComputer Science
Understanding conditionalsthrough conditional structures
A tutorial
Gabriele Kern-Isberner, Christian EichhornAugust 20, 2012
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 1/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?
1 Motivation
2 Some formal preliminaries
3 Conditional Structures
4 Reasoning with conditional structures
5 Properties of structural inference
6 Going beyond structural inference – c-representations
7 Conditional structures and probabilities
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 2/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
1 Motivation
2 Some formal preliminaries
3 Conditional Structures
4 Reasoning with conditional structures
5 Properties of structural inference
6 Going beyond structural inference – c-representations
7 Conditional structures and probabilities
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 3/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
Conditionals
A conditional (B|A) is understood as the uncertain rule“If A, then usually B”.
A probabilistic conditional (B|A)[x] is understood as the rule“If A, then the probability that B holds is expected to be x”.
Example (flying birds)Let b = “bird”, f = “(being able to) fly” , x ∈ [0, 1]
(f |b) “Birds usually fly.”(f |b) [0.9] “Birds fly with a probability of 90 %.”
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 4/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
What are conditionals used for?
Conditionals are rules with exceptions:(Usually,) Philosophers are wise.(However, usually,) Philosophers from the StrangeUniversityare stupid.(But,) Philosopher WiseMan from the StrangeUniversity isvery wise.
A more up-to-date and (maybe very) relevant example:(Usually,) European countries are economically strong.. . .
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 5/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
Why conditionals?
Indeed, some (most?) of the most complicated problems makeuse of conditionals, and it can be of crucial importance to evaluatesituations with respect to such conditionals, e.g., for comparingsituations where the Euro crashes to those where it does not.
But conditionals are also important for our everyday lives:Usually, for buying food I go to the supermarket. But for buyingmeat, I go to the butcher’s, except when it is late in the evening.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 6/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
Why is classical logic not enough, . . .
Classical logic just knows true and false, not plausible vs.implausible, most vs. some, etc..
So, if you say
All philosphers are wise.philosopher ⇒ 1 wise
∀x philosopher(x) ⇒ wise(x),
you really have to mean that – there is no way to allow exceptions!
1⇒ means material implication, i.e., A⇒ B ≡ ¬A ∨ B.Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 7/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
. . . and what makes conditionals so special?
A conditional (B|A) focusses on cases where the premise A isfulfilled but does not say anything about cases when A does nothold.
It leaves semantical room for modelling acceptance in case itsconfirmation A ∧ B is more plausible than its refutationA ∧ ¬B.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 8/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
What are the fallacies of conditional reasoning? I
Conditional reasoning sometimes appears to be fallacious sincemost inference rules (like syllogism, modus ponens, contrapositionetc.) do not hold for conditionals:
Example (Syllogism, classical)A⇒ B, B ⇒ C
A⇒ Cpenguins ⇒ birds, birds ⇒ animals
penguins ⇒ animalsGiven this, we evaluate the statement/situation A ∧ B ∧ C to bemore compliant with reality than A ∧ B ∧ ¬C .
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 9/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?What conditionals are, and why we should care about them
What are the fallacies of conditional reasoning? II
Example (Syllogism, conditionals)However,
penguins are (usually) birds, birds (usually) flypenguins (usually) fly
does not hold!
On the basis of the (uncertain) knowledge about penguins, birds,and flying, how can we evaluate the situationpenguin ∧ bird ∧ ¬flying to be more plausible thanpenguin ∧ bird ∧ flying?
N.B.: The fallacies are not due to the conditionals but to theirimproper usage in classical schemata.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 10/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
1 Motivation
2 Some formal preliminariesLogic and ConditionalsKnowledge bases
3 Conditional Structures
4 Reasoning with conditional structures
5 Properties of structural inference
6 Going beyond structural inference – c-representations
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 11/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
Logic and possible worlds
We use a propositional alphabet Σ generating the propositionallanguage L with the connectors ∧,∨,¬ in the usual way.
AbbreviationsA ∧ B ≡ AB ¬A ≡ A
Possible worlds ω ∈ Ω correspond to interpretations and aredefined in the usual way.
Example
For Σ = a, b, c, Ω =
abc, abc, abc, ab c, abc, abc, a bc, a b c
.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 12/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
Evaluation of conditionalsA conditional (B|A) is understood as the uncertain rule“If A, then usually B”.
Example (flying birds)Let b = “bird”, f = “(being able to) fly”,
(f |b) “Birds usually fly.”
J(B|A)Kω =
1 if ω |= AB (ω satisfies A and B) (verification)0 if ω |= AB (ω satisfies A and not B) (falsification)u if ω |= A (ω satisfies not A)
Definition (Language of conditionals)The language of conditionals is defined as (L | L)
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 13/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
Evaluation of conditionals Examples
Let Σ = p, b, f .
J(b|p)Kpbf = 1 J(b|p)Kpbf = 1 J(b|p)Kpbf = 0 J(b|p)Kpb f = 0J(b|p)Kpbf = u J(b|p)Kpbf = u J(b|p)Kp bf = u J(b|p)Kp b f = u
J(f |b)Kpbf = 1 J(f |b)Kpbf = 0 J(f |b)Kpbf = u J(f |b)Kpb f = uJ(f |b)Kpbf = 1 J(f |b)Kpbf = 0 J(f |b)Kp bf = u J(f |b)Kp b f = u
J(f |p)Kpbf = 0 J(f |p)Kpbf = 1 J(f |p)Kpbf = 0 J(f |p)Kpb f = 1J(f |p)Kpbf = u J(f |p)Kpbf = u J(f |p)Kp bf = u J(f |p)Kp b f = u
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 14/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
Knowledge base
A knowledge base is a (here: finite) set of conditionals
∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L).
Idea:∆ shall represent (all of) the reasoner’s knowledge.Knowlege is “rule-based”.Reasoning shall be (solely) based on ∆.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 15/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
Knowledge base Simple penguin example
The knowlege base ∆ =
(f |b), (f |p), (b|p)
shall represent ourknowledge about penguins, birds and the capability of flying. Weknow, that. . .
(f |b) . . . birds usually fly.(f |p) . . . penguins usually don’t fly.(b|p) . . . penguins are usually birds.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 16/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Logic and Conditionals Knowledge bases
Motivation Examples of exceptions
Penguins are usually Birds (?)(b|p)
Penguins do not fly (?)(f |p)
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 17/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
1 Motivation
2 Some formal preliminaries
3 Conditional StructuresBasic ideasThe formal machineryThe complete definition
4 Reasoning with conditional structures
5 Properties of structural inference
6 Going beyond structural inference – c-representations
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 18/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
Conditional structures Basic ideas
Let ∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L).
The conditional structure of a world ω shall represent theeffect of each conditional in ∆ to ω.This effect shall be independent of the context in which aconditional is applied, i.e., it should not matter which otherconditionals in ∆ are applied or not.This effect shall also be independent of the order in whichthe conditionals are applied to ω.Conditional structures should be comparable to serve as basefor a preference relation on worlds; this preference relationshall express which world is more plausible than another.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 19/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
Introducing symbols a+, a−
Let ∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L). For each conditional(Bi |Ai) ∈ ∆ we define two abstract symbols, a+
i and a−i such thata+
i is linked to the verification AiBi of (Bi |Ai).a−i is linked to the falsification AiBi of (Bi |Ai).
We express this by the following function2:
Definition (σi for each (Bi |Ai))
σi : Ω→ a+i ,a−i , 1 σi(ω) =
a+
i iff ω |= AiBia−i iff ω |= AiBi1 iff ω |= Ai
2“iff” means “if and only if”Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 20/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
ExamplesLet ∆ =
(f |b)︸ ︷︷ ︸
1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
.
σ1(pbf ) = a+1 σ1(pbf ) = a−1 σ1(pbf ) = 1 σ1(pb f ) = 1
σ1(pbf ) = a+1 σ1(pbf ) = a−1 σ1(p bf ) = 1 σ1(p b f ) = 1
σ2(pbf ) = a−2 σ2(pbf ) = a+2 σ2(pbf ) = a−2 σ2(pb f ) = a+
2
σ2(pbf ) = 1 σ2(pbf ) = 1 σ2(p bf ) = 1 σ2(p b f ) = 1
σ3(pbf ) = a+3 σ3(pbf ) = a+
3 σ3(pbf ) = a−3 σ3(pb f ) = a−3σ3(pbf ) = 1 σ3(pbf ) = 1 σ3(p bf ) = 1 σ3(p b f ) = 1
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 21/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
Free abelian3 group
In order to be able to “calculate” with the a−1 ,a+1 , we define
(F∆, ·) as the free abelian group for∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L) with
a−1 ,a+1 , . . . ,a−n ,a+
n
as a basis,
· resp. juxtaposition as multiplication, and1 as the neutral element
F∆ is simply the set of all products of the a−i ,a+i without
cancellations, i.e., no effect of one conditional, either positive ornegative, can be simulated by the effect of another conditional.
3“abelian” means “commutative”Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 22/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
Conditional structureConditional structure shall be defined as the combination of alleffects of all conditionals in ∆:
Definition (conditional structure σ(ω))Let ∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L). The conditionalstructure of a world ω ∈ Ω is defined asa
σ(ω) =n∏
i=1σi(ω) =
∏ω|=AiBi
a+i ·
∏ω|=AiBi
a−i
a∏ is the product symbol, i.e.,∏n
i=1 ai = a1 · . . . · an
Of cause every σ(ω) ∈ F; due to the properties of an abeliangroup, the order of the symbols a+/−
i is irrelevant, and factors 1can be ignored.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 23/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Basic ideas The formal machinery The complete definition
Example for conditional structures
Let ∆ =
(f |b)︸ ︷︷ ︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
.
σ(p b f ) = a+1 a−2 a+
3 σ(p b f ) = a+1 · 1 · 1
σ(p b f ) = a−1 a+2 a+
3 σ(p b f ) = a−1 · 1 · 1σ(p b f ) = 1 · a−2 a−3 σ(p b f ) = 1 · 1 · 11σ(p b f ) = 1 · a+
2 a−3 σ(p b f ) = 1 · 1 · 11
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 24/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
2 Some formal preliminaries
3 Conditional Structures
4 Reasoning with conditional structuresAn idea of preferencePreference relationInference relation
5 Properties of structural inference
6 Going beyond structural inference – c-representations
7 Conditional structures and probabilities
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 25/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Conditional structures Review of basic ideas
Let ∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L).
X The conditional structure of a world ω shall represent theeffect of each conditional in ∆ to ω.
X This effect shall be independent of the context in which aconditional is applied, i.e., it should not matter which otherconditionals in ∆ are applied or not.
X This effect shall also be independent of the order in whichthe conditionals are applied to ω.
? Conditional structures should be comparable to serve as basefor a preference relation on worlds; this preference relationshall express which world is more plausible than another (forevaluating situations).
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 26/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Structural preference Basic ideas
Reminder: A conditional (B|A) is interpreted as the defeasible rule“if A then usually B”.
A world falsifying a conditional defies the associated rule.Defying rules is bad.Ceteris paribus, we want to prefer worlds that do not defy agiven rule to those worlds that do.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 27/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Structural preference The definition
Let ∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L), ω, ω′ possible worlds
A conditional structure σ(ω) is preferred to a conditional structureσ(ω′) (written as σ(ω) ≺σ σ(ω′)) iff
σ(ω′) falsifies every conditional falsified by σ(ω) andσ(ω′) falsifies at least one conditional more than σ(ω).
Definition (σ(ω) ≺σ σ(ω′))
σ(ω) ≺σ σ(ω′) iffσi(ω) = a−i implies σi(ω′) = a−i for all 1 ≤ i ≤ nthere is an i such that σi(ω) ∈
a+
i , 1
and σi(ω′) = a−i
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 28/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Example for structural preference
∆ =
(f |b)︸ ︷︷ ︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
simple penguins knowledge base.
σ(p b f ) = a+1 a−2 a+
3 σ(p b f ) = a−1 a+2 a+
3 σ(p b f ) = a−2 a−3 σ(p b f ) = a+2 a−3
σ(p b f ) = a+1 σ(p b f ) = a−1 σ(p b f ) = 1 σ(p b f ) = 1
E.g., we obtain the following preferences
σ(pb f ) = a+2 a−3 ≺σ a−2 a−3 = σ(pbf )
σ(pbf ) = a+1 ≺σ a+
1 a−2 a+3 = σ(pbf )
σ(pbf ) = 1 ≺σ a+2 a−3 = σ(pb f )
σ(p b f ) = 1 ≺σ a−2 a−3 = σ(pbf )
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 29/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Preference relation on worlds
The preference relation on conditional structures induces apreference relation on worlds:
Definition (Preference relation on worlds)We prefer a world ω to a world ω′ iff the conditional structure of ωis preferred to the conditional structure of ω′, i.e.
ω <σ ω′ iff σ(ω) ≺σ σ(ω′).
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 30/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Inference relation based on conditional structures
Now, we use this preference relation for inference, i.e., to entail(new) plausible knowledge:
Let A,B ∈ L. We structurally infer B from A (A|∼σ∆B) iff forevery world that falsifies (B|A) there is a preferred world thatverifies (B|A). Formally:
Definition (Structural inference by preferential semantics, A|∼σ∆B)
A|∼σ∆B iff for all ω′ |= AB there is an ω |= AB with ω <σ ω′.
In this case, the conditional (B|A) is accepted by |∼σ∆.
Structural inference follows the lines of preferential entailment[Makinson 1994].
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 31/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?An idea of preference Preference relation Inference relation
Examples for structural inference
Let ∆ =
(f |b)︸ ︷︷ ︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
.
σ(p b f ) = a+1 a−2 a+
3 σ(p b f ) = a−1 a+2 a+
3 σ(p b f ) = a−2 a−3 σ(p b f ) = a+2 a−3
σ(p b f ) = a+1 σ(p b f ) = a−1 σ(p b f ) = 1 σ(p b f ) = 1
E.g., we may infer:pb|∼σ∆f since pb f <σ pbfbf |∼σ∆p since pbf <σ pbfb|∼σ∆f since pbf <σ pbf and pbf <σ pbf
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 32/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
2 Some formal preliminaries
3 Conditional Structures
4 Reasoning with conditional structures
5 Properties of structural inferenceHandling of irrelevant informationLimitations of structural inference
6 Going beyond structural inference – c-representations
7 Conditional structures and probabilities
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 33/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Irrelevant information Example
Which influence has information that is obviously irrelevant since itis not mentioned in the knowledge base?
We enlarge our alphabet with r for being red to Σ′ = p, b, f , r.
ω σ(ω) ω σ(ω) ω σ(ω) ω σ(ω)p b f r a+
1 a−2 a+3 p b f r a−2 a−3 p b f r a+
1 p b f r 1p b f r a+
1 a−2 a+3 p b f r a−2 a−3 p b f r a+
1 p b f r 1p b f r a−1 a+
2 a+3 p b f r a+
2 a−3 p b f r a−1 p b f r 1p b f r a−1 a+
2 a+3 p b f r a+
2 a−3 p b f r a−1 p b f r 1
Obviously, r (being true or false) has no influence on theconditional structures of worlds.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 34/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Irrelevant information
Let’s look at inferences – two possibilities:Can the additional information r invalidate inferences, e.g.,we could conclude b|∼σ∆f , but can we also conclude that redbirds fly – br |∼σ∆f ?Can the additional information r induce new inferences, e.g.,are we now able to conclude that birds are red – b|∼σ∆r?
N.B.: The first problem touches monotony but |∼σ∆ is known tobe nonmonotonic. However, one of the most crucial questions ofnonmonotonic reasoning is when exactly should we expect amonotonic behaviour even for nonmonotonic inference relations ?
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 35/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Irrelevant information Example, continued
Are red birds capable of flying? Formally: does br |∼σ∆f hold?
ω σ(ω) ω σ(ω) ω σ(ω) ω σ(ω)p b f r a+
1 a−2 a+3 p b f r a−2 a−3 p b f r a+
1 p b f r 1p b f r a+
1 a−2 a+3 p b f r a−2 a−3 p b f r a+
1 p b f r 1p b f r a−1 a+
2 a+3 p b f r a+
2 a−3 p b f r a−1 p b f r 1p b f r a−1 a+
2 a+3 p b f r a+
2 a−3 p b f r a−1 p b f r 1
σ(p b f r) = a+1 ≺σ a−1 a+
2 a+3 = σ(p b f r) X
σ(p b f r) = a+1 ≺σ a−1 = σ(p b f r) X
br |∼σ∆f holds, red birds can (also) fly.
So, information r proves to be truely irrelevant (as expected).
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 36/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Irrelevant information Example, continued
Are birds usually red? Formally: does b|∼σ∆r hold?ω σ(ω) ω σ(ω) ω σ(ω) ω σ(ω)
p b f r a+1 a−2 a+
3 p b f r a−2 a−3 p b f r a+1 p b f r 1
p b f r a+1 a−2 a+
3 p b f r a−2 a−3 p b f r a+1 p b f r 1
p b f r a−1 a+2 a+
3 p b f r a+2 a−3 p b f r a−1 p b f r 1
p b f r a−1 a+2 a+
3 p b f r a+2 a−3 p b f r a−1 p b f r 1
σ(p b f r) = a+1 ≺σ a+
1 a−2 a+3 = σ(p b f r) X
σ(p b f r) = a+1 ≺σ a−1 a+
2 a+3 = σ(p b f r) X
σ(p b f r) = a+1 ≺σ a−1 = σ(p b f r) X
∅ = ≺σ a+1 = σ(p b f r)
Does b|∼σ∆r hold? – NO; for the same reasons b 6|∼σ∆r , too, so:Are birds usually red? Unknown X.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 37/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Limitations of (purely) structural inference
∆ =
(f |b)︸ ︷︷ ︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
simple penguin knowledge base
σ(p b f ) = a+1 a−2 a+
3 σ(p b f ) = a−1 a+2 a+
3 σ(p b f ) = a−2 a−3 σ(p b f ) = a+2 a−3
σ(p b f ) = a+1 σ(p b f ) = a−1 σ(p b f ) = 1 σ(p b f ) = 1
<σ is no total ordering: e.g. pbf and pbf are incomparable⇒ We are not able to conclude p 6|∼σ∆f !!
I.e. A|∼σ∆B does not hold for all (B|A) ∈ ∆ !!
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 38/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Drowning problem (example)
The drowning problem addresses the problem that exceptionalsubclasses (like penguins) are treated as exceptions throughout,i.e., they may not inherit other typical properties of the superclass.In our simple penguin example, we add the rule that birds usuallyhave wings (w|b) to ∆,
so let ∆′ =
(f |b)︸ ︷︷ ︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
, (w|b)︸ ︷︷ ︸4
.
The question now is:
Can we conclude that (also) penguins have wings?
N.B.: This addresses transitivity resp. the validity of syllogism.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 39/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Drowning problem (example, continued)
ω σ(ω) ω σ(ω) ω σ(ω) ω σ(ω)p b f w a+
1 a−2 a+3 a+
4 p b f w a−2 a−3 p b f w a+1 a+
4 p b f w 1p b f w a+
1 a−2 a+3 a−4 p b f w a−2 a−3 p b f w a+
1 a−4 p b f w 1p b f w a−1 a+
2 a+3 a+
4 p b f w a+2 a−3 p b f w a−1 a+
4 p b f w 1p b f w a−1 a+
2 a+3 a−4 p b f w a+
2 a−3 p b f w a−1 a−4 p b f w 1
Do penguins have wings?
σ(p b f w) = a+1 a−2 a+
3 a+4 ≺σ a+
1 a−2 a+3 a−4 = σ(p b f w) X
σ(p b f w) = a−1 a+2 a+
3 a+4 ≺σ a−1 a+
2 a+3 a−4 = σ(p b f w) X
σ(p b f w) = a+2 a−3 ≺σ a−2 a−3 = σ(p b f w) X
∅ ≺σ a+2 a−3 = σ(p b f w)
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 40/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Drowning problem (example, continued 2)
Why are there no worlds ω |= pw with ω <σ p b f w?Remember: σ(p b f w) = a+
2 a−3 .
ω |= pw why ω 6<σ p b f wp b f w a−2 vs. a+
2p b f w a−1 vs. 1 (= σ1(p b f w))p b f w a−2 a−3 >σ a+
2 a−3p b f w a+
2 a−3 6<σ a+2 a−3
Do penguins have wings?
unknown
So, penguins do not inherit the flying-property of their superclass“birds”.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 41/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Handling of irrelevant information Limitations of structural inference
Drowning problem vs. transitivity
Maybe structural preference is too weak to handle transitiveconclusions by chaining rules?We extend our penguin-example by the conditional (a|f ) – flyingobjects are (usually) airborne:
∆′′ =
(f |b), (f |p), (b|p), (w|b), (a|f ).
Here, we find:b|∼σ∆′′ a,
i.e., (f |b), (a|f ) can be chained successfully.N.B.: This does not involve penguins explicitly!
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 42/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
3 Conditional Structures
4 Reasoning with conditional structures
5 Properties of structural inference
6 Going beyond structural inference – c-representationsMotivationThe link to OCFsC-representations solve existing problems
7 Conditional structures and probabilities
8 Conclusion
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 43/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
Motivation
Conditional structures and structural inference prove to beadequate to capture formally the essence of conditionals. However,problems arise because we are not able to compare falsifications ofdifferent conditionals.
In order to overcome the limitations of structural inference we needto weigh the “severeness of falsification” of conditionals, encodingthe idea that the less plausible an exception to a conditional is, thehigher should be the penalty for falsifying the conditional.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 44/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
Motivation
Example(f |b): exceptions are easy to imagine, non-flying birds are e.g.penguins, dodos, kiwis, ostriches,. . .(b|p): penguins which aren’t birds are quite far-fetched(f |p): flying penguins seem to be very implausible
. . . but we saw such exceptions . . .
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 45/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
Ranking Functions
Conditional structures can be easily combined with OCFs:An Ordinal Conditional Function (OCF) or ranking function κ[Spohn 1988] is a function that assigns a degree ofdisbelief/implausibility to any world ω ∈ Ω.
Definition (κ)κ := Ω→ N∞0 such that:
κ−1(0) 6= ∅κ(A) = min
ω|=Aκ(ω)
κ(B|A) = κ(AB)− κ(A)
Example (ranked flyers)
κ(ω) = 0
κ(ω) = 1
κ(ω) = 2
κ(ω) = 4
p bf p b f p b f
pbf p bf
pbf pb f
pb f
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 46/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
C-representationsA c-representation of ∆ is an OCF with ranks calculated on thebasis of conditional structures [GKI 2001].
Definition (c-representation)
κc∆ : Ω→ N∞0 κc
∆(ω) =∑
1≤i≤nω|=AiB i
κ−i
(i.e., a−i ∼ κ−i , a+i ∼ 0), with κ−i chosen in such a way that
κc∆ |= ∆, i.e.,
κ−i > minω|=AiBi
∑
ω|=AjBji 6=j
κ−j
− minω|=AiBi
∑
ω|=AjBji 6=j
κ−j
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 47/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
c-representations Example: minimal c-representation
c-representation of ∆ =
(f |b)︸︷︷︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
(κ−1 = 1, κ−2 = κ−3 = 2)
ωfalsified
κc∆(ω)conditionals
p b f 2 κ−2 = 2p b f 1 κ−1 = 1p b f 2, 3 κ−2 + κ−3 = 4p b f 3 κ−3 = 2p b f 1 κ−1 = 1p b f — 0p b f — 0p b f — 0
κc∆(ω)
0
1
2
4
p bf p b f p b f
pbf p bf
pbf pb f
pb f
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 48/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
Inference with c-representations
Definition (Inference relation |∼c∆)
A|∼c∆ B iff. κc
∆(AB) < κc∆(AB ) iff. κc
∆ |= (B|A)
Example (are not flying penguins birds in c-representation?)
pf |∼c∆ b
κ(pbf ) = 1 < 2 = κ(pb f )
0
1
2
4
p bf p b f p b f
pbf p bf
pbf pb f
pb f
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 49/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?Motivation The link to OCFs C-representations solve existing problems
Drowning problem (example)
We add the rule that birds usually have wings (w|b) to ∆, so let
∆ =
(f |b)︸︷︷︸1
, (f |p)︸ ︷︷ ︸2
, (b|p)︸ ︷︷ ︸3
, (w|b)︸ ︷︷ ︸4
, (κ−1 = 1, κ−2 = 2, κ−3 = 2, κ−4 = 1)
ω κc∆(ω) ω κc
∆(ω) ω κc∆(ω) ω κc
∆(ω)p b f w κ−2 = 2 p b f w κ−2 + κ−3 = 4 p b f w 0 p b f w 0p b f w κ−2 + κ−4 = 1 p b f w κ−2 + κ−3 = 4 p b f w κ−4 = 1 p b f w 0p b f w κ−1 = 1 p b f w κ−3 = 2 p b f w κ−1 = 2 p b f w 0p b f w κ−1 + κ−4 = 2 p b f w κ−3 = 2 p b f w κ−1 + κ−4 = 2 p b f w 0
Do penguins have wings? Yes, because
κ(pw) = minω|=pw
κc∆(ω) = 1 < 2 = min
ω|=pwκc
∆(ω) = κ(pw)
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 50/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?
3 Conditional Structures
4 Reasoning with conditional structures
5 Properties of structural inference
6 Going beyond structural inference – c-representations
7 Conditional structures and probabilities
8 Conclusion
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 51/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?
Maximum Entropy
MaxEnt Principle: Maximise indeterminacy (i.e. entropy)
H (P) = −∑ω∈Ω
P(ω) · log2 P(ω)
of a probability distribution P givenR = (B1|A1)[x1], . . . , (Bn |An)[xn ],i.e., solve the optimisation problem
(arg) maxP|=R
H (P) = −∑ω∈Ω
P(ω) · log2 P(ω)
to get the probabilistic model P of R which adds as littleinformation as possible.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 52/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?
Conditional structures and maximum entropy
Lagrangre-function of the (above) optimisation problem
−→ ME(R)(ω) = α0 ·∏
1≤i≤nω|=AiBi
α1−xii ·
∏1≤i≤nω|=AiBi
α−xii
Compare this to conditional structures:
σ(ω) =∏
1≤i≤nω|=AiBi
a+i ·
∏1≤i≤nω|=AiBi
a−i
−→ ME(R) is a probabilistic c-representation of R !
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 53/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?
Conclusion
Conditional structuresMake the effect of conditionals in a knowledge base explicitand transparentStructural inference and inductive reasoning from knowledgebasesHowever, have limitations since falsifications of differentconditionals cannot be handled
C-representations are based on conditional structures:Allow to associate conditionals with “severeness offalsification”Solve drowning problemComply with high-standard inference properties (system P, Retc.)Idea can also be used for belief revision
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 54/58
Department ofComputer Science
Motivation Preliminaries Conditional Structures Reasoning with C.S. Inference-Properties C-Representations Probabilities?
Bibliography
Gabriele Kern-Isberner. Conditionals in Nonmonotonic Reasoning and Belief Revision – ConsideringConditionals as Agents, volume 2087. Springer, 2001.
David Makinson. General patterns in nonmonotonic reasoning. In Dov M. Gabbay, C. J. Hogger, and J. A.Robinson, editors, Handbook of logic in artificial intelligence and logic programming, volume 3, pages35–110. Oxford University Press, Inc., New York, NY, USA, 1994. ISBN 0-19-853747-6.
Wolfgang Spohn. Ordinal conditional functions: A dynamic theory of epistemic states. Springer, August31, 1988. ISBN 9027726345. 105–134 pp.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 54/58
Department ofComputer Science
Group theory
(Mathematical) Groups
A Group (G, ∗) consists of a set of group-elements G and a binaryoperation ∗ satisfying the following axioms:
Associativity: For all a, b, c ∈ G it is a ∗ (b ∗ c) = (a ∗ b) ∗ c.
Neutral element:There is an e ∈ G such that for all a ∈ G: a ∗ e = a = e ∗ a.
Inverse element:For all a ∈ G there is an b ∈ G such that a ∗ b = e = b ∗ a
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 55/58
Department ofComputer Science
Group theory
(Mathematical) Groups Examples
The set of integers Z and addition (+) build a group:
Associativity: 2 + (3 + 4) = (2 + 3) + 4Neutral Element (0): 5 + 0 = 5 = 0 + 5Inverse Element: 2 + (−2) = 0 = (−2) + 2
The set of rational numbers Q and multiplication (·) build a group:
Associativity: 2 · (3 · 4) = (2 · 3) · 4Neutral Element (1): 5 · 1 = 5 = 1 · 5Inverse Element: 2 · 1/2 = 1 = 1/2 · 2
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 56/58
Department ofComputer Science
Group theory
Further properties of groups
Definition (Abelian groups)A group (G, ∗) is called abelian iff ∗ is commutative, i.e. for alla, b ∈ G it is a ∗ b = b ∗ a.
Definition (Free abelian groups)A abelian group (G, ∗) is called a free abelian group iff the grouphas a “basis”, i.e. each element of G can be written as linearcombination of the basis (base elements are also called generators).
Example (Free abelian groups)(Z,+) is a free abelian group with base 1.(Q, ·) is a free abelian group with the prime numbers as base.
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 57/58
Department ofComputer Science
Group theory
Free abelian group
We define (F, ·)∆ as the free abelian group for∆ = (B1|A1), . . . , (Bn |An) ⊆ (L | L) with
a−1 ,a
+1 , . . . ,a−n ,a+
n
as base,
· (“usual” multiplication ) as binary relation and1 as neutral element(F is the set of all linear combinations of the base).
Kern-Isberner / Eichhorn Conditional Structures August 20, 2012 58/58