Constraint Satisfaction Problems

Constraint satisfaction problems (CSPs)

• Standard search problem: state is a "black box“ – any data

structure that supports successor function and goal test
– state is defined by variables Xi with values from domain Di
– goal test is a set of constraints specifying allowable combinations
of values for subsets of variables

• Simple example of a formal representation language

• Allows useful general-purpose algorithms with more
power than standard search algorithms
• CSP:
 Example: Map-Coloring

• Variables WA, NT, Q, NSW, V, SA, T

• Domains Di = {red,green,blue}
• Constraints: adjacent regions must have different colors
• e.g., WA ≠ NT, or (WA,NT) in {(red,green),(red,blue),
(green,red), (green,blue),(blue,red),(blue,green)}
 Example: Map-Coloring

• Solutions are complete and consistent assignments

• e.g., WA = red, NT = green, Q = red, NSW =
green,V = red,SA = blue,T = green
•
 Constraint graph
• Binary CSP: each constraint relates two variables
• Constraint graph: nodes are variables, arcs are constraints
 Varieties of CSPs
• Discrete variables
– finite domains:
• n variables, domain size d  O(dn) complete assignments
• e.g., Boolean CSPs, incl. Boolean satisfiability (NP-complete)
– infinite domains:
• integers, strings, etc.
• e.g., job scheduling, variables are start/end days for each job
• need a constraint language, e.g., StartJob1 + 5 ≤ StartJob3

• Continuous variables
– e.g., start/end times for Hubble Space Telescope observations
– linear constraints solvable in polynomial time by LP
 Varieties of constraints
• Unary constraints involve a single variable,
– e.g., SA ≠ green

• Binary constraints involve pairs of variables,

– e.g., SA ≠ WA

• Higher-order constraints involve 3 or more

variables,
– e.g., cryptarithmetic column constraints
 Backtracking search
• Variable assignments are commutative, i.e.,
[ WA = red then NT = green ] same as [ NT = green then WA = red ]
• Depth-first search for CSPs with single-variable assignments
is called backtracking search
• Can solve n-queens for n ≈ 25
• => Only need to consider assignments to a
single variable at each node
Backtracking example
Backtracking example
Backtracking example
Backtracking example
 Improving backtracking efficiency

• General-purpose methods can give huge

gains in speed:
– Which variable should be assigned next?
– In what order should its values be tried?
– Can we detect inevitable failure early?
 Most constrained variable
• Most constrained variable:
choose the variable with the fewest legal values

• a.k.a. minimum remaining values (MRV)

heuristic
–
 Most constraining variable
• A good idea is to use it as a tie-breaker
among most constrained variables
• Most constraining variable:
– choose the variable with the most constraints on
remaining variables
–
 Least constraining value
• Given a variable to assign, choose the least
constraining value:
– the one that rules out the fewest values in the
remaining variables

• Combining these heuristics makes 1000 queens

feasible
–
 Forward checking
• Idea:
– Keep track of remaining legal values for unassigned variables
– Terminate search when any variable has no legal values
–
 Forward checking
• Idea:
– Keep track of remaining legal values for unassigned variables
– Terminate search when any variable has no legal values
–
 Forward checking
• Idea:
– Keep track of remaining legal values for unassigned variables
– Terminate search when any variable has no legal values
–
 Forward checking
• Idea:
– Keep track of remaining legal values for unassigned variables
– Terminate search when any variable has no legal values
–
 Constraint propagation
• Forward checking propagates information from assigned to
unassigned variables, but doesn't provide early detection for all
failures:

• NT and SA cannot both be blue!

• Constraint propagation algorithms repeatedly enforce constraints
locally…
•
 Arc consistency
• Simplest form of propagation makes each arc consistent
• X Y is consistent iff
for every value x of X there is some allowed y

–
•
 Arc consistency
• Simplest form of propagation makes each arc consistent
• X Y is consistent iff
for every value x of X there is some allowed y

–
•
 Arc consistency
• Simplest form of propagation makes each arc consistent
• X Y is consistent iff
for every value x of X there is some allowed y

• If X loses a value, neighbors of X need to be rechecked

–
•
 Arc consistency
• Simplest form of propagation makes each arc consistent
• X Y is consistent iff
for every value x of X there is some allowed y

• If X loses a value, neighbors of X need to be rechecked

• Arc consistency detects failure earlier than forward checking
• Can be run as a preprocessor or after each assignment

•
 Arc consistency algorithm AC-3

• Time complexity: O(#constraints |domain|3)

Checking consistency of an arc is O(|domain|2)
 k-consistency
• A CSP is k-consistent if, for any set of k-1 variables, and for any consistent
assignment to those variables, a consistent value can always be assigned to
any kth variable
• 1-consistency is node consistency
• 2-consistency is arc consistency
• For binary constraint networks, 3-consistency is the same as path consistency
• Getting k-consistency requires time and space exponential in k
• Strong k-consistency means k’-consistency for all k’ from 1 to k
– Once strong k-consistency for k=#variables has been obtained, solution
can be constructed trivially
• Tradeoff between propagation and branching
• Practitioners usually use 2-consistency and less commonly 3-consistency
 Other techniques for CSPs
• Global constraints
– E.g., Alldiff
– E.g., Atmost(10,P1,P2,P3), i.e., sum of the 3 vars ≤ 10
– Special propagation algorithms
• Bounds propagation
– E.g., number of people on two flight D1 = [0, 165] and D2 = [0, 385]
– Constraint that the total number of people has to be at least 420
– Propagating bounds constraints yields D1 = [35, 165] and D2 = [255, 385]
• …

• Symmetry breaking
Structured CSPs
Tree-structured CSPs
Algorithm for tree-structured CSPs
Nearly tree-structured CSPs

(Finding the minimum cutset is NP-complete.)

 Tree decomposition
• Every variable in original
problem must appear in at least
one subproblem
• If two variables are connected
in the original problem, they
must appear together (along
with the constraint) in at least
one subproblem
• If a variable occurs in two
subproblems in the tree, it must
appear in every subproblem on
the path that connects the two

• Algorithm: solve for all solutions of each subproblem. Then, use the tree-
structured algorithm, treating the subproblem solutions as variables for those
subproblems.
• O(ndw+1) where w is the treewidth (= one less than size of largest subproblem)
• Finding a tree decomposition of smallest treewidth is NP-complete, but good
heuristic methods exists
An example CSP application:
satisfiability
 Davis-Putnam-Logemann-Loveland
(DPLL) tree search algorithm
clause
E.g. for 3SAT
p1
? p s.t. (p1p3p4)  (p1p2p3)  …
p2T F
T F
p3
p4 Complete
Backtrack when some clause becomes empty

Unit propagation (for variable & value ordering): if some clause

only has one literal left, assign that variable the value that satisfies
the clause (never need to check the other branch)

Boolean Constraint Propagation (BCP): Iteratively apply unit

propagation until there is no unit clause available
 A helpful observation for the
DPLL procedure
P1  P2  …  Pn  Q (Horn)
is equivalent to
(P1  P2  …  Pn)  Q (Horn)
is equivalent to
P1  P2  …  Pn  Q (Horn clause)
Thrm. If a propositional theory consists only of Horn clauses
(i.e., clauses that have at most one non-negated variable) and
unit propagation does not result in an explicit contradiction
(i.e., Pi and Pi for some Pi), then the theory is satisfiable.
Proof. On the next page.
…so, Davis-Putnam algorithm does not need to branch on
variables which only occur in Horn clauses
 Proof of the thrm

Assume the theory is Horn, and that unit propagation has completed
(without contradiction). We can remove all the clauses that were
satisfied by the assignments that unit propagation made. From the
unsatisfied clauses, we remove the variables that were assigned values by
unit propagation. The remaining theory has the following two types of
clauses that contain unassigned variables only:
P1  P2  …  Pn  Q and
P1  P2  …  Pn
Each remaining clause has at least two variables (otherwise unit
propagation would have applied to the clause). Therefore, each
remaining clause has at least one negated variable. Therefore, we can
satisfy all remaining clauses by assigning each remaining variable to
False.
Variable ordering heuristic for DPLL [Crawford & Auton AAAI-93]
Heuristic: Pick a non-negated variable that occurs in a non-Horn (more than
1 non-negated variable) clause with a minimal number of non-negated
variables.
Motivation: This is effectively a “most constrained first” heuristic if we view
each non-Horn clause as a “variable” that has to be satisfied by setting
one of its non-negated variables to True. In that view, the branching
factor is the number of non-negated variables the clause contains.
Q: Why is branching constrained to non-negated variables?
A: We can ignore any negated variables in the non-Horn clauses because
– whenever any one of the non-negated variables is set to True the clause becomes
redundant (satisfied), and
– whenever all but one of the non-negated variables is set to False the clause becomes
Horn.
Variable ordering heuristics can make several orders of magnitude
difference in speed.
Constraint learning aka nogood learning aka clause learning
used by state-of-the-art SAT solvers (and CSP more generally)

Conflict graph
• Nodes are literals
• Number in parens shows the search tree level
where that node got decided or implied

• Cut 2 gives the first-unique-implication-point (i.e., 1 UIP on reason side) constraint

(v2 or –v4 or –v8 or v17 or -v19). That schemes performs well in practice.
Any cut would give a valid clause. Which cuts should we use? Should we delete some?
• The learned clauses apply to all other parts of the tree as well.
 Conflict-directed backjumping

x7=0

Failure-driven assertion (not a branching decision):

Learned clause is a unit clause under this path, so
BCP automatically sets x7=0.

x2=0 • Then backjump to the decision level of x3=1,

keeping x3=1 (for now), and
forcing the implied fact x7=0 for that x3=1 branch
• WHAT’S THE POINT? A: No need to just backtrack to x2
Classic readings on conflict-directed backjumping,
clause learning, and heuristics for SAT
• “GRASP: A Search Algorithm for Propositional Satisfiability”,
Marques-Silva & Sakallah, IEEE Trans. Computers, C-48, 5:506-
521,1999. (Conference version 1996.)
• (“Using CSP look-back techniques to solve real world SAT instances”,
Bayardo & Schrag, Proc. AAAI, pp. 203-208, 1997)
• “Chaff: Engineering an Efficient SAT Solver”, Moskewicz, Madigan,
Zhao, Zhang & Malik, 2001 (
www.princeton.edu/~chaff/publication/DAC2001v56.pdf)
• “BerkMin: A Fast and Robust Sat-Solver”, Goldberg & Novikov, Proc.
DATE 2002, pp. 142-149
• See also slides at
http://www.princeton.edu/~sharad/CMUSATSeminar.pdf
More on conflict-directed backjumping (CBJ)
• These are for general CSPs, not SAT specifically:
• Read Section 6.3.3. of Russell & Norvig for an easy description of
conflict-directed backjumping for general CSP
• “Conflict-directed backjumping revisited” by Chen and van Beek,
Journal of AI Research, 14, 53-81, 2001:
– As the level of local consistency checking (lookahead) is increased, CBJ
becomes less helpful
• A dynamic variable ordering exists that makes CBJ redundant
– Nevertheless, adding CBJ to backtracking search that maintains generalized
arc consistency leads to orders of magnitude speed improvement
experimentally
• “Generalized NoGoods in CSPs” by Katsirelos & Bacchus, National
Conference on Artificial Intelligence (AAAI-2005) pages 390-396, 2005.
– This paper generalizes the notion of nogoods, and shows that nogood learning
(then) can speed up (even non-SAT) CSPs significantly
 Random restarts
• Sometimes it makes sense to keep restarting the
CSP/SAT algorithm, using randomization in
variable ordering
– Avoids the very long run times of unlucky variable
ordering
– On many problems, yields faster algorithms
– Clauses learned can be carried over across restarts
– Experiments suggest it does not help on optimization
problems (e.g., [Sandholm et al. IJCAI-01, Management
Science 2006])
Phase transitions in CSPs
 “Order parameter” for 3SAT
[Mitchell, Selman, Levesque AAAI-92]

•  = #clauses / # variables
• This predicts
– satisfiability
– hardness of finding a model
How would you capitalize on the
phase transition in an algorithm?
Generality of the order parameter 

• The results seem quite general across model

finding algorithms
• Other constraint satisfaction problems have
order parameters as well
…but the complexity peak does not occur
(at least not in the same place) under all
ways of generating SAT instances
Iterative refinement algorithms
for SAT
 GSAT [Selman, Levesque, Mitchell AAAI-92]
(= a local search algorithm for model finding)

Avg. total flips Incomplete (unless restart a lot)

2000 50 variables, 215 3SAT clauses
1600
1200 Greediness is not essential as long
800 as climbs and sideways moves are
400
100 200
max-climbs preferred over downward moves.
 BREAKOUT algorithm [Morris AAAI-93]
Initialize all variables Pi randomly
UNTIL current state is a solution
IF current state is not a local minimum
THEN make any local change that reduces the total cost
(i.e. flip one Pi)
ELSE increase weights of all unsatisfied clause by one

Incomplete, but very efficient on large (easy) satisfiable problems.

Reason for incompleteness: the cost increase of the current local

optimum spills over to other solutions because they share
unsatisfied clauses.