on computing all abductive explanations from a propositional horn theory
DESCRIPTION
Joint work with Thomas Eiter (Technische Universitat Wien). On Computing all Abductive Explanations from a Propositional Horn Theory. Kaz Makino (Graduate School of Engineering Science, Osaka Univ.). Outline. 3 reasoning mechanisms Abduction from Horn theories - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/1.jpg)
On Computing all Abductive Explanations from a Propositional Horn Theory
Kaz Makino (Graduate School of Engineering Science, Osaka Univ.)
Joint work with Thomas Eiter (Technische Universitat Wien)
..
![Page 2: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/2.jpg)
Outline
1. 3 reasoning mechanisms
2. Abduction from Horn theories
3. Generating abductive explanations from Horn theories
4. Model-based representation for Horn theories
![Page 3: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/3.jpg)
Deduction: fact knowledge base ?
Induction: fact ? observation
Abduction: ? knowledge base observation
3 Reasoning Mechanisms
Fact: battery is down
knowledge: if the battery is down, the car will not start ……………………………
Deduction
The car will not start
![Page 4: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/4.jpg)
Deduction: fact knowledge base ?
Induction: fact ? observation
Abduction: ? knowledge base observation
3 Reasoning Mechanisms
Induction
Rule: if the battery is down, the car will not start
Fact: battery is down
Observation: The car will not start
![Page 5: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/5.jpg)
Deduction: fact knowledge base ?
Induction: fact ? observation
Abduction: ? knowledge base observation
3 Reasoning Mechanisms
Abduction
Fact: battery is down
knowledge: if the battery is down, the car will not start ……………………………
Observation: The car will not start
![Page 6: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/6.jpg)
Abduction (formulated by C.S. Peirce ’31-’58)
Basis for
Truth Maintenance Systems (TMS, ATMS) Clause management Systems (CMS) Diagnosis Database Update, ...
Widely used in Computer Science and AI
![Page 7: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/7.jpg)
Propositional Horn Knowledge Base
Propositional variables:
CNF (conjunctive normal form): E.g.,
Horn CNF: at most 1 positive literal in each clause
Knowledge
![Page 8: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/8.jpg)
Propositional Horn Knowledge Base
Horn CNF: at most 1 positive literal in each clause
Horn rule:
Horn clause:
(antecedent/consequent: may be empty)
Core language in AI and logic programming
![Page 9: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/9.jpg)
Horn CNF representation is not uniqueE.g.,
![Page 10: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/10.jpg)
Explanations
: a propositional variable: a Horn CNF
(1)
(2) s.t.
(1)
(2) is satisfiable
An explanation for from : a minimal set s.t.
: implies for all
![Page 11: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/11.jpg)
Explanations
: a propositional variable: a Horn CNF
(1)
(2) is satisfiable
An explanation for from : a minimal set s.t.
Abduction: ? knowledge base observation
![Page 12: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/12.jpg)
E.g., Explanations for from
: trivial explanation for
(1)
(2) is satisfiable
Well-known: Finding a nontrivial explanation is poly. time. Finding an explanation is NP-hard. (Selman & Levesque '90)
![Page 13: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/13.jpg)
Abduction: ? knowledge base observation
: if the battery is down, the car will not start if the gas tank is empty, the car will not start ……………………………
: The car will not start
: {battery is down} , , …
Car diagnosis
{gas tank is empty}
1. Generate all possible explanations 2. Find a real one from them
![Page 14: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/14.jpg)
Can we generate all (poly. many) explanations efficiently ?
Eiter & Makino (2002) disproved it
Conjecture by Selman & Levesque ('90)
Generating explanations is NP-hard, even if there are only few explanations overall.
Note: Exponentially many explanations might exist.
explanations
![Page 15: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/15.jpg)
Complexity of generating problem
start … … halt
P delay :
Output P :Slow
Fast
Incremental P :
![Page 16: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/16.jpg)
Prime implicate of
Ex.
・
・
![Page 17: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/17.jpg)
(1)
(2) is satisfiable
Explanation : minimal s.t.
nontrivial explanation for
prime implicate containing
Explanations and prime implicates
![Page 18: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/18.jpg)
How to generate all prime implicates
resolvent
E.g., : resolvent of
![Page 19: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/19.jpg)
Procedure ResolutionInput: A CNF representing Output: All prime implicates ofStep 1:Step 2: Repeat (s) simplification and (r) resolution.
(s) Remove from if s.t.
(r) Add a resolvent of two clauses in
Step 3: Output all clauses in
Ex.
![Page 20: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/20.jpg)
Prop. [Blake ('37), Brown ('68), Quine ('55), Samson-mills ('54)]
Resolution procedure generates all prime implicates.
Prop. There is no output P algorithm for generating all prime implicates, unless P=NP.
Prop. Even if is Horn, resolution procedure may require exponential time.
![Page 21: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/21.jpg)
Modification
Procedure ResolutionInput: A CNF representing Output: All prime implicates ofStep 1:Step 2: Repeat (s) simplification and (r) resolution.
(s) Remove from if s.t.
(r) Add a resolvent of two clauses in
Step 3: Output all clauses in
(1) Input resolution:(2) Add a prime implicate s.t.(3) Output in (r) immediately, if is new.
![Page 22: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/22.jpg)
Th. [Boros, Crama, Hammer ('90)]If is Horn, then input-resolution procedure generates all prime implicates in incremental P time.
nontrivial explanation for
prime implicate containing
![Page 23: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/23.jpg)
Modification
Procedure ResolutionInput: A CNF representing Output: All prime implicates ofStep 1:Step 2: Repeat (s) simplification and (r) resolution.
(s) Remove from if s.t.
(r) Add a resolvent of two clauses in
Step 3: Output all clauses in
(1) Input resolution:(2) Add a prime implicate s.t.(3) Output in (r) immediately, if is new.(4)
![Page 24: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/24.jpg)
Th. [Eiter, Makino (2002, 2003)]All explanations for from a Horn CNF can be computed with P delay.
Cor. P many explanations for from a Horn CNF can be computed in (input) P time.
Sketch:
![Page 25: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/25.jpg)
nontrivial explanation for a negative literal
prime implicate containing
Explanations for a negative literal
![Page 26: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/26.jpg)
Modification
Procedure ResolutionInput: A CNF representing Output: All prime implicates ofStep 1:Step 2: Repeat (s) simplification and (r) resolution.
(s) Remove from if s.t.
(r) Add a resolvent of two clauses in
Step 3: Output all clauses in
(1) Input resolution:(2) Add a prime implicate s.t.(3) Output in (r) immediately, if is new.(4)
![Page 27: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/27.jpg)
All explanations for from an acyclic Horn CNF can be computed in incremental P time.
Th. [Eiter, Makino (2003)]
Explanations for a negative literal
Th. [Eiter, Makino (2003)]There exists no output P algorithm for generating all explanations for from a Horn CNF, unless P=NP.
Prop. [Eiter, Makino (2003)]Our resolution procedure does not generate all explanations for from a Horn CNF.
![Page 28: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/28.jpg)
Summary
Knowlege
Horn CNF P delay no output P
Acyclic Horn CNF P delay incremental P
Characteristic set MDual MDual
Explanations query query
Knowlege
Horn CNF coNPc coNPc
Acyclic Horn CNF coNPc coNPc
Characteristic set MDual MDual
Explanations query query
![Page 29: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/29.jpg)
Model-based reasoningDeduction
E.g.,
All models satisfy
does not satisfy
large inefficient
![Page 30: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/30.jpg)
[P. N. Johnson-Laird ('83)].
Humans typically argue by just looking at some examples.
Of course, incorrect !
Some models satisfy
conclude otherwise,
![Page 31: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/31.jpg)
intersection of
E.g.,
Horn CNF: syntactic characterization
Prop. [McKinsey ('43)]a Horn function
is closed under
Semantic, model theoretic characterization
![Page 32: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/32.jpg)
Model-based representation of a Horn function
for maximal modelRelational Database: generating set
E.g.,
Intersection closure
Characteristic set
![Page 33: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/33.jpg)
Horn CNFs
Finding a nontrivial explanation is poly. time. Finding an explanation is NP-hard.
Given Characteristic set
Dedution: poly. time
Finding a nontrivial explanation is poly. time. Finding an explanation is poly. time.
![Page 34: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/34.jpg)
Summary
Knowlege
Horn CNF P delay no output P
Acyclic Horn CNF P delay incremental P
Characteristic set MDual MDual
Explanations query query
Knowlege
Horn CNF coNPc coNPc
Acyclic Horn CNF coNPc coNPc
Characteristic set MDual MDual
Explanations query query
![Page 35: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/35.jpg)
Monotone Dualization
Output : Prime DNF of
Ex.
Input: A CNF of (a monotone function)
![Page 36: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/36.jpg)
Monotone Dualization
Output : Prime DNF of
Input: A CNF of (a monotone function)
Many P equivalent problems
![Page 37: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/37.jpg)
Polynomial ? OPEN
Bioch, Boros, Crama, Domingo, Eiter, Elbassioni, Fredman, Gaur, Gogic,Gottlob, Gunopulos, Gurvich, Hammer, Ibaraki, Johnson, Kameda, Kavvadias, Khachiyan, Khardon, Kogan, Krishnamurti, Lawler, Lenstra, Lovasz, Mannila, Mishra, Papadimitriou, Pitt, Rinnoy Kan, Sideri, Stavropoulos, Tamaki, Toinonen, Uno, Yannakakis, ...
´
![Page 38: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/38.jpg)
Polynomial ? OPEN
茨木俊秀 . 単調論理関数の同定問題とその複雑さ , 離散構造とアルゴリズム III, 室田一雄(編) 近代科学社 ,pp. 1--33, 1994. Toshihide Ibaraki
Johnson. Open and closed problems in NP-completeness. Lecture given at the International School of Mathematics ``G. Stampacchia'': Summer School ``NP-Completeness:The First 20 Years'', Erice, Italy, June 20-27, 1991.
Lovasz. Combinatorial optimization: Some problems
and trends, DIMACS Technical Report 92-53, 1992.
![Page 39: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/39.jpg)
Papadimitriou. NP-completeness: A retrospective,In: Proc. 24th International Colloquium on Automata, Languages and Programming (ICALP), pp.2--6, Springer LNCS 1256, 1997.
Mannila. Local and Global Methods in Data Mining: Basic Techniques and Open Problems In: Proc. 29th ICALP, pp.57--68, Springer LNCS 2380, 2002.
Eiter, Gottlob. Hypergraph Transversal Computation and Related Problems in Logic and AI, In: Proc. European Conference on Logics in Artificial Intelligence (JELIA),
pp. 549-564, Springer LNCS 2224, 2002
![Page 40: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/40.jpg)
Best known
[Fredman, Khachiyan, 94]
time where
[Eiter, Gottlob, Makino, 02]
guessed bits
![Page 41: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/41.jpg)
Summary
Knowlege
Horn CNF P delay no output P
Acyclic Horn CNF P delay incremental P
Characteristic set MDual MDual
Explanations query query
Knowlege
Horn CNF coNPc coNPc
Acyclic Horn CNF coNPc coNPc
Characteristic set MDual MDual
Explanations query query
![Page 42: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/42.jpg)
Explanations
Knowlege general DNF CNF pos Horn general pos neg general
Horn CNF coNPc nOP coNPc Pd coNPc coNPc Pd nOP nOP
coNPc nOP coNPc nOP MD nOP MD nOP nOP
Explanations query
clause term
Knowlege general DNF CNF pos Horn general pos neg general
Horn CNF coNPc coNPc coNPc
coNPc nOP MD nOP MD coNPc coNPc
queryclause term
nOP: no Output P, Pd: P delay, MD: Monotone Dual
![Page 43: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/43.jpg)
1. Abductive Inference
2. Monotone Dualization
3. Horn Transformation
4. Vertex Enumeration
Open Problems
![Page 44: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/44.jpg)
![Page 45: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/45.jpg)
Conclusion
Generating abductive explanations from Horn CNFs
Practical side
High order logic, non-Horn case.
![Page 46: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/46.jpg)
![Page 47: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/47.jpg)
![Page 48: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/48.jpg)
Modification
Procedure ResolutionInput: A CNF representing Output: All prime implicates ofStep 1:Step 2: Repeat (s) simplification and (r) resolution.
(s) Remove from if s.t.
(r) Add a resolvent of two clauses in
Step 3: Output all clauses in
(1) Input resolution:(2) Add a prime implicate s.t.(3) Output in (r) immediately, if is new.(4)
![Page 49: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/49.jpg)
Sketch: All explanations can be generated by the input resolution procedure
No negative clause
if is definite Horn: CNF consisting of all prime implicates
: CNF consisting of all prime implicates generated so far
・
prime
: irredundant ( )Note
![Page 50: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/50.jpg)
: CNF consisting of all prime implicates
: CNF consisting of all prime implicates generated so far
・
![Page 51: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/51.jpg)
![Page 52: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/52.jpg)
![Page 53: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/53.jpg)
a contradiction.
Proof.
Horn clause
From
or
![Page 54: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/54.jpg)
![Page 55: On Computing all Abductive Explanations from a Propositional Horn Theory](https://reader036.vdocuments.site/reader036/viewer/2022062408/568134ea550346895d9c2274/html5/thumbnails/55.jpg)