Logical reduction of metarules Andrew Cropper & Sophie Tourret - - PowerPoint PPT Presentation

logical reduction of metarules
SMART_READER_LITE
LIVE PREVIEW

Logical reduction of metarules Andrew Cropper & Sophie Tourret - - PowerPoint PPT Presentation

Logical reduction of metarules Andrew Cropper & Sophie Tourret ILP Examples Learner Computer program Background knowledge Mode declarations Progol ILASP Aleph XHAIL TILDE Metarules Metagol MIL-Hex ILP ProPPR Clint


slide-1
SLIDE 1

Logical reduction of metarules

Andrew Cropper & Sophie Tourret

slide-2
SLIDE 2

Examples Background knowledge Learner Computer program ILP

slide-3
SLIDE 3

Mode declarations Progol ILASP Aleph XHAIL TILDE …

slide-4
SLIDE 4

Metagol MIL-Hex ∂ILP ProPPR Clint MOBAL … (almost all neural-ILP approaches) Metarules

slide-5
SLIDE 5

Metarules

P(A,B) ← Q(A,C),R(C,B)

slide-6
SLIDE 6

Metarules

P(A,B) ← Q(A),R(A,B)

slide-7
SLIDE 7

Metarules

P(A,B) ← Q(A),R(A,B)

slide-8
SLIDE 8

Input % background parent(alice,bob). parent(bob,charlie). % example grandparent(alice,charlie). % metarule P(A,B) ← Q(A,C),R(C,B)

slide-9
SLIDE 9

Output % subs Subs = {P\grandparent,Q\parent,R\parent} % program grandparent(A,B) ← parent(A,C), parent(C,B)

slide-10
SLIDE 10

Where do we get metarules from?

slide-11
SLIDE 11

Completeness Cannot learn grandparent/2 with only P(X)←Q(X) Efficiency More metarules = larger hypothesis space

slide-12
SLIDE 12

Idea: find logically minimal sets

slide-13
SLIDE 13

The clause C is entailment redundant in the clausal theory T ∪ {C} when T ⊨ C

Entailment redundant metarules [Cropper and Muggleton, ILP14]

slide-14
SLIDE 14

C1 = p(A,B)←q(A,B) C2 = p(A,B)←q(A,B),r(A) C3 = p(A,B)←q(A,B),r(A),s(B,C) Entailment redundancy

slide-15
SLIDE 15

{C1} ⊨ {C2,C3} C1 = p(A,B)←q(A,B) C2 = p(A,B)←q(A,B),r(A) C3 = p(A,B)←q(A,B),r(A),s(B,C) Entailment redundancy

slide-16
SLIDE 16

P(A,B)←Q(A,B) P(A,B)←Q(B,A) P(A,B)←Q(A,C),R(B,C) P(A,B)←Q(A,C),R(C,B) P(A,B)←Q(B,A),R(A,B) P(A,B)←Q(B,A),R(B,A) P(A,B)←Q(B,C),R(A,C) P(A,B)←Q(B,C),R(C,A) P(A,B)←Q(C,A),R(B,C) P(A,B)←Q(C,A),R(C,B) P(A,B)←Q(C,B),R(A,C) P(A,B)←Q(C,B),R(C,A)

?

Entailment reduction

slide-17
SLIDE 17

P(A,B)←Q(B,A) P(A,B)←Q(A,C),R(C,B) Entailment reduction P(A,B)←Q(A,B) P(A,B)←Q(B,A) P(A,B)←Q(A,C),R(B,C) P(A,B)←Q(A,C),R(C,B) P(A,B)←Q(B,A),R(A,B) P(A,B)←Q(B,A),R(B,A) P(A,B)←Q(B,C),R(A,C) P(A,B)←Q(B,C),R(C,A) P(A,B)←Q(C,A),R(B,C) P(A,B)←Q(C,A),R(C,B) P(A,B)←Q(C,B),R(A,C) P(A,B)←Q(C,B),R(C,A)

slide-18
SLIDE 18

C1 = P(A,B)←Q(A,B) C2 = P(A,B)←Q(A,B),R(A) C3 = P(A,B)←Q(A,B),R(A,B) C4 = P(A,B)←Q(A,B),R(A,B),S(A,B)

  • Entailment reduction problem
slide-19
SLIDE 19

{C1} ⊨ {C2,C3,C4} C1 = P(A,B)←Q(A,B) C2 = P(A,B)←Q(A,B),R(A) C3 = P(A,B)←Q(A,B),R(A,B) C4 = P(A,B)←Q(A,B),R(A,B),S(A,B)

  • Entailment reduction problem
slide-20
SLIDE 20

father(A,B) ← parent(A,B), male(A) ✖

  • Entailment reduction problem

C1 = P(A,B)←Q(A,B) C2 = P(A,B)←Q(A,B),R(A) C3 = P(A,B)←Q(A,B),R(A,B) C4 = P(A,B)←Q(A,B),R(A,B),S(A,B) {C1} ⊨ {C2,C3,C4}

slide-21
SLIDE 21

The clause C is derivationally redundant in the theory T ∪ {C} when T ⊢ C

Derivation redundancy [Cropper and Tourret, ILP18, JELIA19]

slide-22
SLIDE 22

The clause C is derivationally redundant in the theory T ∪ {C} when T ⊢ C

SLD-resolution in this work

Derivation redundancy

slide-23
SLIDE 23

C1 = P(A,B)←Q(A,B) C2 = P(A,B)←Q(A,B),R(A) C3 = P(A,B)←Q(A,B),R(A,B) C4 = P(A,B)←Q(A,B),R(A,B),S(A,B) Derivation redundancy

slide-24
SLIDE 24

{C1,C2,C3} ⊢ {C4} C1 = P(A,B)←Q(A,B) C2 = P(A,B)←Q(A,B),R(A) C3 = P(A,B)←Q(A,B),R(A,B) C4 = P(A,B)←Q(A,B),R(A,B),S(A,B) father(A,B) ← parent(A,B), male(A) ✔ Derivation redundancy

slide-25
SLIDE 25

MLJ Paper

We compare subsumption, entailment, and derivation reduction We theoretically show whether infinite fragments of metarules can be logically reduced to finite sets We run the reduction algorithms on finite sets of metarules to identify minimal sets We experimentally compare the learning performance

  • f Metagol when supplied with reduced sets of

metarule

slide-26
SLIDE 26
  • Q. Can we reduce M to a fragment with only two

body literals?

  • Q. Can we reduce M to a fragment with finitely many

body literals?

  • Q. If M has a finite reduction, what is that fragment?

Theoretical questions

slide-27
SLIDE 27

Idea

  • 1. Generate big sets of metarules
  • 2. Run the reduction algorithms on the sets
  • 3. Study the results.
slide-28
SLIDE 28

Ham maximum arity a maximum body literals m

slide-29
SLIDE 29

Can the connected fragment C2∞ be reduced to two body literals?

slide-30
SLIDE 30

Can the connected fragment C2∞ be reduced to two body literals?

slide-31
SLIDE 31

C2∞ cannot be derivationally reduced to C22 Can the connected fragment C2∞ be reduced to two body literals?

slide-32
SLIDE 32

C2∞ cannot be derivationally reduced to a finite fragment! Can the connected fragment C2∞ be reduced to two body literals?

slide-33
SLIDE 33

Derivation reduction of connected fragment

slide-34
SLIDE 34

Why not?

slide-35
SLIDE 35

Can the Datalog D2∞ fragment be reduced to two body literals?

slide-36
SLIDE 36

Can the Datalog D2∞ fragment be reduced to two body literals? D2∞ cannot be derivationally reduced to a finite fragment

slide-37
SLIDE 37

Reduction summary

slide-38
SLIDE 38

Does it matter? Is there any difference in learning performance when using different reduced sets of metarules?

slide-39
SLIDE 39

Trains

slide-40
SLIDE 40

Accuracies

slide-41
SLIDE 41

Learning times

slide-42
SLIDE 42

String transformations

slide-43
SLIDE 43

String transformations

slide-44
SLIDE 44

Inducing game rules

slide-45
SLIDE 45

Inducing game rules

slide-46
SLIDE 46

Conclusions New form of logical reduction Negative theoretical results (especially for MIL) Little impact on practical performance

slide-47
SLIDE 47

Todo Overcome negative theoretical result Expand results to higher-arities Identify domain-specific sets of metarules Identify optimal sets of metarule Logically reduce BK