towards theoretical frameworks for comparing constraint satisfaction models and algorithms peter van...

45
Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus November 2001

Upload: gavin-baker

Post on 28-Mar-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

Towards theoretical frameworks for comparing constraint satisfaction models and algorithms

Peter van Beek, University of WaterlooCP 2001 · Paphos, CyprusNovember 2001

Page 2: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

2

Acknowledgements

Joint work with:

Fahiem Bacchus

Xinguang Chen

Grzegorz Kondrak

Toby Walsh

Page 3: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

3

Constraint programming methodology

Model problem·specify in terms of constraints on acceptable solutions

·define/choose constraint model: variables, domains, constraints

Solve model·define/choose algorithm·define/choose heuristics

Verify and analyze solution

Page 4: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

4

Previous work

Many models, algorithms, and heuristics proposed

Choice of model, algorithm, and heuristic can greatly influence efficiency

(e.g., Nadel, 1990; Ginsberg et al., 1990;

Frost and Dechter, 1994; Tsang et al., 1995 Smith et al., 2000; Beacham et al., 2001)

Page 5: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

5

How to compare?

Empirical studies·benchmark sets·random problems

Theoretical studies·worst case·average case

(e.g. Brown & Purdom, 1981;

Nadel, 1983; Wilf, 1984)

·pair-wise comparisons-partial orders-bounded worse

Performance measures

·nodes visited·constraint checks·CPU time

Page 6: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

6

Part I

Comparing backtracking algorithms

Page 7: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

7

Some backtracking algorithms

Chronological Backtracking (BT)

Backjumping (BJ)(Gaschnig, 1978)

Conflict-Directed Backjumping (CBJ)(Prosser, 1993)

Forward Checking (FC)(Haralick & Elliott, 1980; McGregor, 1979)

Maintaining arc consistency (MAC)(Haralick & Elliott, 1980; McGregor, 1979; Sabin & Freuder, 1994)

Page 8: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

8

Example: 4-queens

4

3

2

1

x1 x2 x3 x4

xi xj and

| xi - xj | | i - j |

Page 9: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

9

Search tree for 4-queens

x1

x2x3

x4

1 2 3 4

(1,1,1,1) (4,4,4,4)(2,4,1,3) (3,1,4,2)

Page 10: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

10

4

3

2

1

{x1 1}

Backjumping

Q

Q{x1 1, x2

3}

x1

x2

x1

x2

{x1 1, x2 3 , x4

2}

Q

x3

x1 x3 x4x2

Page 11: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

11

“New” algorithms: BJ1, …, BJn

BJk :·allowed to backjump at most k times from a leaf·after that it must chronologically backtrack

Special cases:·BJ1 equivalent to BJ·BJn equivalent to CBJ

Page 12: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

12

k-consistency

Given ·any k-1 distinct variables,·any consistent assignment to those variables,·any additional variable x

there exists a consistent assignment to x

4

3

2

1

x1 x2 x3 x4

2-consistent ?

Q

Q1-consistent ?

3-consistent ?strongly 2-

consistent

yes

yes

no

Page 13: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

13

Maintaining consistency

{ x1 1, x2

4 }

4

3

2

1

x1 x2 x3 x4

Q

Qx3 x4

|x3 - x4| 1

Binary constraints

Unary constraintsx3 1, 3, 4

x4 1, 2, 4

Induced CSP

Page 14: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

14

Maintaining consistency

4

3

2

1

x1 x2 x3 x4

Q

Q

{ x1 1, x2

4 }

x3 x4

|x3 - x4| 1

Binary constraints

Unary constraintsx3 1, 3, 4

x4 1, 2, 4

FC: strong 1-consistency

Page 15: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

15

Maintaining consistency

Q

Q

{ x1 1, x2

4 }

x3 x4

|x3 - x4| 1

Binary constraints

Unary constraintsx3 1, 3, 4

x4 1, 2, 4

4

3

2

1

x1 x2 x3 x4

MAC: strong 2-consistency

Q

Q

Page 16: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

16

“New” algorithms: MC1, …, MCn

MCk :·maintains strong k-consistency on induced CSP at each node in backtrack tree

Special cases:·MC1 equivalent to FC·MC2 equivalent to MAC for binary CSPs

Page 17: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

17

Comparing backtracking algorithms

Proposal:for all CSP models for all variable orderings

algorithm a1 algorithm a2

Desirable features:·still holds if use best possible CSP model for a2

·still holds if use best possible variable ordering for a2

Page 18: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

18

Theoretical methodology

Using local consistency concepts:·formulate necessary and sufficient conditions for a node to be visited by a backtracking algorithm

Using conditions:·construct partial orders of the algorithms

Page 19: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

19

4

3

2

1

x1 x2 x3 x4

{x1 1}

{x1 1, x2

4}?

Example sufficient condition: FC

If a node is consistent and its parent is strongly 1-consistent, then FC visits the node

Q

Q

Page 20: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

20

Example necessary condition: MAC-CBJ

If MAC-CBJ visits a node, then it is consistent and its parent is strongly 2-consistent

4

3

2

1

x1 x2 x3 x4

{x1 1}

{x1 1, x2

4}?

Q

Page 21: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

21

Example conclusion: MAC-CBJ vs FC

FC visits p

MAC-CBJ visits p

p is consistent and parent(p) strongly 1-consistent

p is consistent and parent(p) strongly 2-consistent

Page 22: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

22

Partial order for nodes visited

MCk+1-CBJ

MCk-CBJ

MC1-CBJ

MCn-CBJ

MCk+1

MCk

MC1

MCn

BJk+1

BJk

BJ1

BJn

BT

...

...

...

...

...

...

less than or equal

Page 23: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

23

Partial order for nodes visited

MCk+1-CBJ

MCk-CBJ

MC1-CBJ

MCn-CBJ

MCk+1

MCk

MC1

MCn

BJk+1

BJk

BJ1

BJn

BT

...

...

...

......

...

incomparable

Page 24: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

24

Partial order for nodes visited

BT

MCk+1-CBJ

MCk-CBJ

MC1-CBJ

MCn-CBJ

MCk+1

MCk

MC1

MCn

BJk+1

BJk

BJ1

BJn

...

...

...

......

...

incomparable

Page 25: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

25

Partial order for nodes visited

MCk+1-CBJ

MCk-CBJ

MC1-CBJ

MCn-CBJ

MCk+1

MCk

MC1

MCn

BJk+1

BJk

BJ1

BJn

BT

...

...

...

...

...

...

incomparable

Page 26: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

26

Partial order for nodes visited

MCk+1-CBJ

MCk-CBJ

MC1-CBJ

MCn-CBJ

MCk+1

MCk

MC1

MCn

CBJBT

...

......

...

Page 27: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

27

Closer look: nodes visited, binary CSPs

MAC-CBJ

FC-CBJ

MAC

FC

CBJ

BJ2

BJ

BT

less than or equal

Page 28: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

28

Closer look: constraint checks, binary CSPs

MAC-CBJ

FC-CBJ

MAC

FC

CBJ

BJ2

BJ

BT

less than or equal

at most O(f) more

O(nd)

O(n2d2)

O(nd)

Page 29: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

29

Experiments: lookahead

MAC vs FC Faster? CommentsMcGregor (1979) FC 3 fasterHaralick & Elliott (1980) FC 3 fasterSabin & Freuder (1994) MAC much betterBacchus & van Run (1995) FC 3 20 fasterBessiere & Regin (1996) MAC much betterLarrosa (2000) both much better

Path vs MAC Faster? CommentsBacchus & van Run (1995) MAC much better

FC

MAC

O(nd)

Path

BT

O(nd)

O(nd)

Page 30: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

30

Example: FC vs MAC

Adapted from J. Ullman (1988); suggested by D. De Haan

Odd-even relation

x y1 21 42 12 33 23 44 14 3

variables: x1, …, xn, n odd

domains: {1, 2, 3, 4}

x1 x2 xn...

FC: 2n+1 - 4

MAC: 4

Page 31: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

31

Increasing lookahead

Problem MC2 MC3 MC4

1 > 3600 262 602 794 40 1403 7 3 34 1065 21 25 34 17 26 > 3600 125 37 58 6 48 17 3 29 163 95 210 87 4 2

Time (sec.) to solve 6 x 6 puzzlesMC4

MC3

MC2

MC1

Page 32: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

32

Experiments: lookback

FC vs FC-CBJ Faster? CommentsProsser (1993) FC-CBJ three times betterFrost & Dechter (1994) FC-CBJ somewhat betterBacchus & van Run (1995) FC-CBJ slightlySmith & Grant (1995) FC-CBJ sometimes much betterBayardo & Schrag (1996, 1997) FC-CBJ much betterLi & Anbulagan (1997) FC as good or better

FC-CBJFC

Page 33: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

33

Experiments: lookahead vs lookback

Increasing the level of consistency decreases effectiveness of CBJ

·Prosser (1993)·Bacchus & van Run (1995) ·Bessiere & Regin (1996)

MAC much better than FC-

CBJ

MAC better than MAC-CBJ

MAC-CBJMAC

FC-CBJFC

O(nd)

CBJBT

O(nd)

“CBJ becomes useless”

· Jussien, Debruyne, Boizumault (2000)MAC-DBT much better than MAC (binary)

·Chen & van Beek (2001)MGAC-CBJ much better than MAC (non-binary)

Page 34: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

34

Theory: lookahead vs lookback

MCk+1-CBJ

MCk-CBJ

MCk+1

MCk

BJk+1

BJk

BJk+1 better only if there exists a (k+1)-level backjump that is not a chronological backtrack

Number of (k+1)-level backjumps number of k-level backjumps

Therefore, as k increases, effectiveness of backjumping decreases

Page 35: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

35

Part II

Comparing CSP models

Page 36: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

36

Conversion to binary

Any non-binary CSP can be converted into one with only binary constraints

·dual graph method (Dechter & Pearl, 1989; Rossi et al. 1989)

·hidden variable method (Peirce, 1933; Rossi et al. 1989; Dechter 1990)

Page 37: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

37

Crossword puzzles

1 2 3

6

4

7

5

8

10

9

20

11

22

12

21

13

17

14

181615

23

19

aaardvarkabackabacusabaftabaloneabandon...

Mona Lisamonarchmonarchymonarda...zymurgyzyrianzythum

Page 38: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

38

Non-binary model (original)

1 2 3

6

4

7

5

8

10

9

20

11

22

12

21

13

17

14

181615

23

19

variables: one for each

unknown letter (cell)

domains: ‘a’, …, ‘z’

constraints: contiguous

letters must form words in dictionary

Page 39: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

39

Dual model (binary)

1 2 3

6

4

7

5

8

10

9

20

11

22

12

21

13

17

14

181615

23

19

variables: one for each unknown word across and down

domains: words from

dictionary

constraints intersecting words must agree on common letter

Page 40: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

40

Hidden model (binary)

1 2 3

6

4

7

5

8

10

9

20

11

22

12

21

13

17

14

181615

23

19

variables: one for each unknown letter (cell) and

one for each unknown word across and down

domains: letters, words

constraints: letter and word

variable must agree

Page 41: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

41

Comparing models

One proposal:for all backtracking algorithms for all variable orderings

model m1 model m2

Objections:·effectiveness of a model is algorithm dependent

·models may contain different variables

Desirable features:·meets objections·still holds if use best possible variable ordering for m2

Another proposal:given backtrack algorithm for all v2 for m2

there exists a v1 for m1 s. t. model m1 model m2

Page 42: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

42

Theoretical methodology

Bound difference in number of nodes visited·compare pruning power of local consistencies of alternative models

(e.g., Debruyne & Bessiere, 1997;

Stergiou & Walsh, 1999; Prosser et al.,

2000; Walsh, 2000)·establish correspondence between nodes

Bound difference in cost at each node:·local consistency·variable ordering

Page 43: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

43

Some relationships

MGAC-dual MGAC-hiddenMGAC-orig

FC-dual FC-hiddenFC-orig

BT-dual BT-hiddenBT-orig

polynomial not polynomial

Page 44: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

44

Part III

Comparing variable ordering heuristics

Future work

Page 45: Towards theoretical frameworks for comparing constraint satisfaction models and algorithms Peter van Beek, University of Waterloo CP 2001 · Paphos, Cyprus

45

Conclusion

Technical:·comparing backtracking algorithms

partial order·comparing CSP models

bounded worse relation on algorithm-model combinations

Big picture:·theory can guide, inform experimentation