1. CSPs Tamara Berg CS 590-133 Artificial Intelligence Many slides throughout the course adapted from Svetlana Lazebnik, Dan Klein, Stuart Russell, Andrew Moore, Percy Liang, Luke Zettlemoyer

2. Review from last class

3. What is search for? Assumptions: single agent, deterministic, fully observable, discrete environment Search for planning – The path to the goal is the important thing – Paths have various costs, depths Search for assignment – Assign values to variables while respecting certain constraints – The goal (complete, consistent assignment) is the important thing

4. Constraint satisfaction problems (CSPs) Definition: – X i is a set of variables {X 1,… X n } – D i is a set of domains {D 1,... D n } one for each variable – C is a set of constraints that specify allowable combinations of values – Solution is a complete, consistent assignment How does this compare to the “generic” tree search formulation? – A more structured representation for states, expressed in a formal representation language – Allows useful general-purpose algorithms with more power than standard search algorithms by eliminating large portions of search space (identifying var/val combinations that violate constraints).

5. Example: Map Coloring Variables: WA, NT, Q, NSW, V, SA, T Domains: {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)}

6. 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

7. Example: Cryptarithmetic Variables: T, W, O, F, U, R X 1, X 2 Domains: {0, 1, 2, …, 9} Constraints: O + O = R + 10 * X 1 W + W + X 1 = U + 10 * X 2 T + T + X 2 = O + 10 * F Alldiff(T, W, O, F, U, R) T ≠ 0, F ≠ 0 X 2 X 1

8. Example: Sudoku Variables: X ij Domains: {1, 2, …, 9} Constraints: Alldiff(X ij in the same unit) X ij

9. Real-World CSPs Assignment problems – e.g., who teaches what class – hard constraint - no prof can teach 2 classes at the same time. – preference constraints – Jan prefers teaching in the morning while Alex prefers teaching in the afternoon (soft constraints/costs). For example an afternoon slot for Jan can cost 2 points whereas a morning slot costs 1.

10. Real-World CSPs Timetable problems – e.g., which class is offered when and where? – Classroom size & enrollment can put hard constraints on where a class can be held – Avoiding simultaneous scheduling of two related classes can be a soft constraint

11. Real-World CSPs Transportation Scheduling – E.g. airline scheduling Flights F 1,F 2 with capacities 165 and 385 – Domains D 1 =[0,165] D 2 =[0,385] – Suppose we have an additional constraint that the two flights must carry 420 people: ie F 1 +F 2 =420 – Then we can reduce the domains to: D 1 =[35,165] D 2 =[255,385] Enforcing Bounds consistency

12. Real-World CSPs Job-shop scheduling – e.g. assembling a car – Variables – X = tasks {Axle f, Axle b, Wheel rf, Wheel lf, Wheel rb, Wheel lb,Nuts rf, Nuts lf, Nuts lb, Nuts rb, Cap rf, Cap lf, Cap rb, CAP lb, Inspect} – Values – D i time task starts – Constraints – one ask must occur before other, a task takes a certain amount of time to complete. e.g. Axle F + 10 ≤ Wheel rf, Wheel rf +1≤Nuts rf, Nutsrf+2≤Cap rf … 4 workers, but only 1 axle tool: – Axle f +10≤Axle b or Axle b +10≤Axle f For each variable, X, except Inspect: – X+d x ≤ Inspect Want total time less than 30 minutes and inspect takes 3 minutes: – D i = {1,2,3,….,27}

13. More examples of CSPs: http://www.csplib.org/ http://www.csplib.org/

14. Types of Constraints Unary – restricts value of a single variable, e.g. SA≠green. Binary – relates two variables, e.g. SA≠NSW Higher order – relates more than two variables, e.g. Between (X,Y,Z) Global – arbitrary number of variables – commonly Alldiff – all variables must have different values. e.g. in sudoku all variables in a unit must satisfy Alldiff. Preference constraint – Red is better than green – Often representable by a cost for each variable assignment Resource constraint – Atmost constraint. e.g. scheduling P 1,P 2,P 3,P 4 denote personnel assigned to each task. Constrain is that you only have ≤ 10 people. So Atmost(10,P 1,P 2,P 3,P 4 )

15. Variations on CSPs Simplest CSPs – discrete, finite domains map coloring, scheduling with time limits, 8-queens Discrete with infinite domain no deadline scheduling Continuous domains scheduling of experiments on Hubble telescope Linear programming problems constraints must be linear equalities or inequalities (can be solved in time polynomial in # vars)

16. Variables: X ij Domains: {1, 2, …, 9} Constraints: Alldiff(X ij in the same unit) X ij Why are constraints useful? They let us eliminate portions of assignment space By identifying var/val combinations that violate constraints ? ?

18. Standard search formulation (incremental) States: – Variables and values assigned so far Initial state: – The empty assignment Action: – Choose any unassigned variable and assign to it a value that does not violate any constraints Fail if no legal assignments Goal test: – The current assignment is complete and satisfies all constraints

19. Standard search formulation (incremental) What is the depth of any solution (assuming n variables)? n (this is good) Given that there are m possible values for any variable, how many paths are there in the search tree? n! · m n (this is bad) How can we reduce the branching factor?

20. Backtracking search In CSP’s, variable assignments are commutative – For example, [WA = red then NT = green] is the same as [NT = green then WA = red] We only need to consider assignments to a single variable at each level (i.e., we fix the order of assignments)

21. Example

25. Backtracking search In CSP’s, variable assignments are commutative – For example, [WA = red then NT = green] is the same as [NT = green then WA = red] We only need to consider assignments to a single variable at each level (i.e., we fix the order of assignments) Depth-first search for CSPs with single-variable assignments is called backtracking search

26. Backtracking search algorithm

29. Which variable should be assigned next? Most constrained variable: – Choose the variable with the fewest legal values – A.k.a. minimum remaining values (MRV) heuristic

30. Which variable should be assigned next? Most constrained variable: – Choose the variable with the fewest legal values – A.k.a. minimum remaining values (MRV) heuristic

31. Which variable should be assigned next? Most constraining variable: – Choose the variable that imposes the most constraints on the remaining variables

32. Which variable should be assigned next? Most constraining variable: – Choose the variable that imposes the most constraints on the remaining variables – Tie-breaker among most constrained variables

33. Backtracking search algorithm

34. Given a variable, in which order should its values be tried? Choose the least constraining value: – The value that rules out the fewest values in the remaining variables

35. Given a variable, in which order should its values be tried? Choose the least constraining value: – The value that rules out the fewest values in the remaining variables Which assignment for Q should we choose?

36. Early detection of failure Apply inference to reduce the space of possible assignments and detect failure early

37. Early detection of failure: Forward checking Keep track of remaining legal values for unassigned variables Terminate search when any variable has no legal values

38. Early detection of failure: Forward checking Keep track of remaining legal values for unassigned variables Terminate search when any variable has no legal values

39. Early detection of failure: Forward checking Keep track of remaining legal values for unassigned variables Terminate search when any variable has no legal values

40. Early detection of failure: Forward checking Keep track of remaining legal values for unassigned variables Terminate search when any variable has no legal values

41. Early detection of failure: Forward checking Keep track of remaining legal values for unassigned variables Terminate search when any variable has no legal values

42. 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 repeatedly enforces constraints locally

43. Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y Arc consistency Consistent!

44. Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y Arc consistency

45. Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y – When checking X  Y, throw out any values of X for which there isn’t an allowed value of Y If X loses a value, all pairs Z  X need to be rechecked Arc consistency

46. Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y – When checking X  Y, throw out any values of X for which there isn’t an allowed value of Y If X loses a value, all pairs Z  X need to be rechecked

47. Arc consistency Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y – When checking X  Y, throw out any values of X for which there isn’t an allowed value of Y If X loses a value, all pairs Z  X need to be rechecked

48. Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y – When checking X  Y, throw out any values of X for which there isn’t an allowed value of Y Arc consistency

49. Simplest form of propagation makes each pair of variables consistent: – X  Y is consistent iff for every value of X there is some allowed value of Y – When checking X  Y, throw out any values of X for which there isn’t an allowed value of Y Arc consistency detects failure earlier than forward checking Can be run before or after each assignment Arc consistency

50. Arc consistency algorithm AC-3

53. Specialized constraint checking Constraints involving an arbitrary number of variables can have special purpose algorithms E.g. Alldiff – Remove any variable in the CSP that has a singleton domain and remove that value from the domains of remaining variables. – Repeat as long as you have singleton variables. – If at any point an empty domain is produced or you have more variables than domain values left - inconsistency!

54. Specialized constraint checking Constraints involving an arbitrary number of variables can have special purpose algorithms E.g. Atmost – Detect inconsistencies by checking sum of min values in current domain. – Can also enforce consistencies by deleting max values of any domain if it’s not consistent with min values in other domains Atmost(10,P 1,P 2,P 3,P 4 ) If domains are {2,3,4,5,6} can delete {5,6} from all domains b/c inconsistent with min domain values

56. Tree-structured CSPs Certain kinds of CSPs can be solved without resorting to backtracking search! Tree-structured CSP: constraint graph does not have any loops Source: P. Abbeel, D. Klein, L. Zettlemoyer

57. Algorithm for tree-structured CSPs Choose one variable as root, order variables from root to leaves such that every node's parent precedes it in the ordering http://cs188ai.wikia.com/wiki/Tree_Structure_CSPs

58. Algorithm for tree-structured CSPs Choose one variable as root, order variables from root to leaves such that every node's parent precedes it in the ordering Backward removal phase: check arc consistency starting from the rightmost node and going backwards http://cs188ai.wikia.com/wiki/Tree_Structure_CSPs

59. Algorithm for tree-structured CSPs Choose one variable as root, order variables from root to leaves such that every node's parent precedes it in the ordering Backward removal phase: check arc consistency starting from the rightmost node and going backwards Forward assignment phase: select an element from the domain of each variable going left to right. We are guaranteed that there will be a valid assignment because each arc is arc consistent http://cs188ai.wikia.com/wiki/Tree_Structure_CSPs

60. Algorithm for tree-structured CSPs Running time is O(nm 2 ) (n is the number of variables, m is the domain size) – We have to check arc consistency once for every node in the graph (every node has one parent), which involves looking at pairs of domain values

61. Nearly tree-structured CSPs Cutset conditioning: find a subset of variables whose removal makes the graph a tree, instantiate that set in all possible ways, prune the domains of the remaining variables and try to solve the resulting tree-structured CSP Source: P. Abbeel, D. Klein, L. Zettlemoyer

62. Algorithm for tree-structured CSPs Running time is O(nm 2 ) (n is the number of variables, m is the domain size) – We have to check arc consistency once for every node in the graph (every node has one parent), which involves looking at pairs of domain values What about backtracking search for general CSPs? – Worst case O(m n ) Can we do better?

63. Computational complexity of CSPs The satisfiability (SAT) problem: The satisfiability (SAT) problem – Given a Boolean formula, is there an assignment of the variables that makes it evaluate to true? SAT is NP-complete (Cook, 1971)NP-complete – NP: class of decision problems for which the “yes” answer can be verified in polynomial time – An NP-complete problem is in NP and every other problem in NP can be efficiently reduced to it – Other NP-complete problems: graph coloring, n-puzzle, generalized sudoku – Open question: is P = NP? Open question: is P = NP?

65. Local search for CSPs Start with “complete” states, i.e., all variables assigned Allow states with unsatisfied constraints Attempt to improve states by reassigning variable values Hill-climbing search: – In each iteration, randomly select any conflicted variable and choose value that violates the fewest constraints – I.e., attempt to greedily minimize total number of violated constraints h = number of conflicts

66. Local search for CSPs Start with “complete” states, i.e., all variables assigned Allow states with unsatisfied constraints Attempt to improve states by reassigning variable values Hill-climbing search: – In each iteration, randomly select any conflicted variable and choose value that violates the fewest constraints – I.e., attempt to greedily minimize total number of violated constraints – Problem: local minima h = 1