Knowledge Representation, Coalgebraically Dirk Pattinson, Imperial - - PowerPoint PPT Presentation

knowledge representation coalgebraically
SMART_READER_LITE
LIVE PREVIEW

Knowledge Representation, Coalgebraically Dirk Pattinson, Imperial - - PowerPoint PPT Presentation

Knowledge Representation, Coalgebraically Dirk Pattinson, Imperial College London (based on joint work with Rajeev Gor, Clemens Kupke and Lutz Schrder) Oxford, October 2010 Back in Tudor England . . . Henry VIII Henry Carey Mary Boleyn


slide-1
SLIDE 1

Knowledge Representation, Coalgebraically

Dirk Pattinson, Imperial College London (based on joint work with Rajeev Gorè, Clemens Kupke and Lutz Schröder) Oxford, October 2010

slide-2
SLIDE 2

Back in Tudor England . . .

Henry VIII Henry Carey Mary Boleyn

“There has been speculation that Mary’s two children, Catherine and Henry, were fathered by Henry, but this has never been proven”

October 30, 2010 1

slide-3
SLIDE 3

What do we know?

20%

  • Mary Boleyn was Henry’s Mistress time between 1520 and 1526 (approx.)
  • suppose that there’s a 20 % chance that Henry Carey is a royal bastard

October 30, 2010 2

slide-4
SLIDE 4

What’s a good model?

Quantitative Uncertainty about offspring:

{ , , , . . . }

0.2 0.2 0.6

{ , , , , . . . } { , , . . . }

Models are probability distributions over sets of successors: W → D(P(W))

October 30, 2010 3

slide-5
SLIDE 5

What’s a good model?

Certainty about amorous affairs: Madge Shelton Elizabeth Blount Models are relations: W → P(W) Combinations of both facets:

W → D(P(W))

  • ffspring

× P(W)

affairs

October 30, 2010 4

slide-6
SLIDE 6

Logical Language

Syntax based on

  • a set P of propositional variables (concepts) like king
  • a set N of nominals (individual variables) like henry
  • a set Λ of modal operators (like “had an affair with”)

Formulas

H(Λ) ∋ A, B ::= x | A∧B | A∨B | ♥(A1, . . . , An) | ¯ ♥(A1, . . . , An) | @iA

where x ∈ P ∪ ¯

P ∪ N ∪ ¯ N, i ∈ N and ♥ ∈ Λ is n-ary.

Interpretation.

  • nominals denote singletons, @i moves evaluation context to point i
  • modalities “scan successors” for relevant properties item M, w |

= A for some w ∈ W — A is satisfiable in M

October 30, 2010 5

slide-7
SLIDE 7

How do we interpret this?

Tudor Example. Λ = {✸} ∪ {Lp | p ∈ Q ∩ [0, 1]} Models are triples M = (W, σ, π) where

σ : W → TW = D(P(W)) × P(W)

and

π : P ∪ N → P(W)

are such that π(n) is a singleton for all n ∈ N. Modalities (where AM = {w ∈ W | M, w |

= A} is the truth-set of A) M, w | = ✸A ⇐ ⇒ σ(w) ∈ {(µ, S) ∈ TW | S ∩ AM = ∅} M, w | = LpA ⇐ ⇒ σ(w) ∈ {(µ, S) ∈ TW | µ({S′ ⊆ W | S′ ∩ AM = ∅}) ≥ p}

Satisfaction Operators and Variables (where x ∈ P ∪ N)

M, w | = @iA ⇐ ⇒ M, π(i) | = A M, w | = x ⇐ ⇒ w ∈ π(x)

October 30, 2010 6

slide-8
SLIDE 8

Enter Coalgebra . . .

Models in the Tudor example: M = (W, σ, π) where

σ : W → TW = P(D(W)) × P(W)

Modalities in the Tudor Example:

M, w | = ✸A ⇐ ⇒ σ(w) ∈ {(µ, S) ∈ TW | S ∩ AM = ∅} M, w | = LpA ⇐ ⇒ σ(w) ∈ {(µ, S) ∈ TW | µ({S′ ⊆ W | S′ ∩ AM = ∅}) ≥ p}

Coalgebraic Models. M = (W, σ, π) where σ : W → TW Coalgebraic Modalities.

M, w | = ♥(A) ⇐ ⇒ σ(w) ∈ ♥W (AM)

and ♥ is a predicate lifting, i.e. a natural family of mappings of type

♥X : P(X) → P(TX)

October 30, 2010 7

slide-9
SLIDE 9

Coalgebraic Setup

Given.

  • a collection Λ of modal operators
  • a functor T : Set → Set inducing T -models (C, γ : C → TC)
  • an interpretation ♥ : P → P ◦ T of every ♥ ∈ Λ
  • Problem. Decide whether A ∈ H(Λ) is satisfiable in Mod(Ξ), for Ξ ⊆ H(Λ).

Moreover, what’s the complexity and what’s a feasible algorithm? Examples.

  • the Tudors (as we’ve seen before)
  • and many more by varying T and Λ (graded, multi-agent, conditional etc.)
  • also quantum phenomena?

October 30, 2010 8

slide-10
SLIDE 10

Enter Tableau Methods . . .

Basic Ingredient. One-Step (Tableau) Rules

♥1A1, . . . , ♥nAn, ¬♠1B1, . . . , ¬♠kBk Σ1 . . . Σk

where Σ1, . . . , Σk ⊆ {±A1, . . . , ±Bk}. Basic Soundness / Completeness If Γ = ♥1A1, . . . , ♥nAn, ¬♠1B1, . . . , ¬♠kBk and τ(Ai), τ(Bj) ⊆ X then

Γ = ∅ ⇐ ⇒ Γi = ∅

for some rule

Γ0 Γ1 ... Γk with Γ0 ⊆ Γ and some 1 ≤ i ≤ n where

  • ♥A = ♥X(τ(A))
  • A = τ(A)
  • Σ1, Σ2 = Σ1 ∩ Σ2

(These rulesets are known for many examples.)

October 30, 2010 9

slide-11
SLIDE 11

Basic Tableaux

Basic Result. Suppose that A ∈ H(Λ) does not contain any nominals or

@-operators. Then A unsatisfiable in Coalg(T) ⇐ ⇒ A has closed tableau

where tableaux are constructed from one-step rules and propositional rules

Γ, A ∧ B Γ, A, B Γ, ¬(A ∧ B) Γ, ¬A Γ, ¬B Γ, A, ¬A

and Coalg(T) is the class of all T -coalgebras (C, γ : C → TC). Proof Idea. Construct satisfying model ‘step-by-step’ by basic completeness

  • Complexity. PSPACE for tractable rule sets

October 30, 2010 10

slide-12
SLIDE 12

Global Assumption Tableaux

Basic Result. Suppose that A ∈ H(Λ) and Θ ⊆ H(Λ) do not contain any nominals or @-operators. Then

A unsatisfiable in Mod(Θ) ⇐ ⇒ A, Θ has closed Θ- tableau

where Θ-tableaux are constructed from propositional and augmented one-step rules

♥1A1, . . . , ♥nAn, ¬♠1B1, . . . , ¬♠kBk Σ1, Θ . . . Σk, Θ

and Mod(Θ) is the class of all T -coalgebras (C, γ : C → TC) validating all formulas in Θ. Proof Idea. Augmented rules force validity of Θ at all constructed states.

  • Complexity. EXPTIME for tractable rule sets.

October 30, 2010 11

slide-13
SLIDE 13

Enter Nominals and Satisfaction Operators . . .

  • Task. For A ∈ H(Λ) and Θ ⊆ H(Λ), decide whether A satisfiable in Mod(Θ).

Complexity Theorist’s Take. That’s easy . . .

  • for each nominal i, guess its theory Ki in the subformulas of A, Θ
  • use Θ-tableau to check satisfiability of A, Θ replacing @iA by ⊤ if A ∈ Ki

and by ⊥, otherwise

  • additionally check satisfiability of Ki.

Proof Idea. Satisfiability entails that Ki as required exist

  • Complexity. EXPTIME as there are only exponentially many possible guesses.

Practitioner’s Complaint. Guessing Ki is ‘too much work’.

October 30, 2010 12

slide-14
SLIDE 14

Coalgebraic Tableaux in Action

Basic Idea. Satisfiability as a game played by ∃ (claiming sat) and ∀

  • every rule application (challenge of ∀) must have a satisfiable conclusion (∃)
  • rule application may uncover @-prefixed formulas that need to be propagated
  • propagated @-formulas can be challenged by ∀

Example 1. ∃ shows satisfiability

✸@ip, ✸(✸@iq ∨ ✸A) @ip, @iq @ip ✸@iq ∨ ✸A i, p, q, @ip, @iq ✸@iq @iq

October 30, 2010 13

slide-15
SLIDE 15

Coalgebraic Tableaux in Action

Basic Idea. Satisfiability as a game played by ∃ (claiming sat) and ∀

  • every rule application (challenge of ∀) must have a satisfiable conclusion (∃)
  • rule application may uncover @-prefixed formulas that need to be propagated
  • propagated @-formulas can be challenged by ∀

Example 2. ∀ shows unsatisfiability

✸@i⊥, ✸A @i⊥, • @i⊥ i, ⊥, @i⊥

  • Note. In both examples, wins were announced without fully unfolding the tableau.

October 30, 2010 14

slide-16
SLIDE 16

Conclusions

  • Complexity. For coalgebraic logics, satisfiability is EXPTIME-decidable over a

(finite) set of global assumptions

  • this has been known before, but we have a new proof
  • Practicability. The decision procedure is purely syntax driven
  • amenable to ‘usual’ optimisations for tableau-bases algorithms
  • room for heuristics
  • Novelty. Even for (description) logics with relational semantics, our algorithm

appears to be new? Still Missing (but not for long): implementation and experiments.

October 30, 2010 15

slide-17
SLIDE 17

Technicalities: Caching Graphs

Basic Entities.

  • Sequents are finite sets of formulas
  • @-constraints are finite sets of @-prefixed formulas possibly containing •

[ • signifies incomplete information ] Expansion Rules

  • for sequents: propositional logic + modal rules (logic specific, but known)
  • for @-constraints:

Υ i, Υ/@i, Υ \ {•}

where i ∈ N(Υ) is a nominal occurring in Υ and Υ/@i = {A | @iA ∈ Υ}. [ Υ is needed as further constraints may be uncovered by modal unfolding ]

October 30, 2010 16

slide-18
SLIDE 18

Expansion and Propagation in a Nutshell

Sequent Expansion →S

  • select unexpanded sequent (X) and create/reuse nodes for all applicable rules
  • create associated @-constraint containing just •
  • mark sequent as unknown (U)

Constraint Expansion →C.

  • select unexpanded constraint (T ) and create/reuse nodes for all applicable rules
  • mark constraint as done (D)

Constraint Propagation →P

  • percolate constraints through the graph using greatest-fixpoint construction

Position Update →U .

  • Update winning positions for ∀ inductively: add sequents that are unsat
  • Update winning positions for ∃ dually (using coinduction)

October 30, 2010 17

slide-19
SLIDE 19

Properties of Ensuing Algorithm

  • Correctness. Given that the modal rules are sound and complete, sequents

marked ‘satisfiable’ are indeed satisfiable (dually for unsat).

  • Proof. If a sequent is marked ‘sat’, player ∃ has a winning strategy in a game played
  • n sequents, which induces a satisfying model.
  • Termination. If modal rules are decidable, every execution terminates and all

sequents will be marked as either sat or unsat.

  • Proof. Determinacy of the associated two-player game.
  • Complexity. If modal rules are EXPTIME decidable, every execution of the

algorithm will terminate in EXPTIME.

  • Proof. The number of sequents that appear in the game is exponential in the size of

the root sequent. Non-Determinism leaves room for heuristics.

October 30, 2010 18