2. Learning to Search Henry Kautz University of Washington joint work with Dimitri Achlioptas, Carla Gomes, Eric Horvitz, Don Patterson, Yongshao Ruan, Bart Selman CORE – MSR, Cornell, UW

3. Speedup Learning Machine learning historically considered Learning to classify objects Learning to search or reason more efficiently Speedup Learning Speedup learning disappeared in mid-90’s Last workshop in 1993 Last thesis 1998 What happened? It failed. It succeeded. Everyone got busy doing something else.

4. It failed. Explanation based learning Examine structure of proof trees Explain why choices were good or bad (wasteful) Generalize to create new control rules At best, mild speedup (50%) Could even degrade performance Underlying search engines very weak Etzioni (1993) – simple static analysis of next-state operators yielded as good performance as EBL

5. It succeeded. EBL without generalization Memoization No-good learning SAT: clause learning Integrates clausal resolution with DPLL Huge win in practice! Clause-learning proofs can be exponentially smaller than best DPLL (tree shaped) proof Chaff (Malik et al 2001) 1,000,000 variable VLSI verification problems

6. Everyone got busy. The something else: reinforcement learning. Learn about the world while acting in the world Don’t reason or classify, just make decisions What isn’t RL?

7. Another path Predictive control of search Learn statistical model of behavior of a problem solver on a problem distribution Use the model as part of a control strategy to improve the future performance of the solver Synthesis of ideas from Phase transition phenomena in problem distributions Decision-theoretic control of reasoning Bayesian modeling

8. Big Picture Problem Instances Solver static features runtime Learning / Analysis Predictive Model dynamic features resource allocation / reformulation control / policy

9. Case Study 1: Beyond 4.25 Problem Instances Solver static features runtime Learning / Analysis Predictive Model

10. Phase transitions & problem hardness Large and growing literature on random problem distributions Peak in problem hardness associated with critical value of some underlying parameter 3-SAT: clause/variable ratio = 4.25 Using measured parameter to predict hardness of a particular instance problematic! Random distribution must be a good model of actual domain of concern Recent progress on more realistic random distributions...

11. Quasigroup Completion Problem (QCP) NP-Complete Has structure is similar to that of real-world problems - tournament scheduling, classroom assignment, fiber optic routing, experiment design,... Can generate hard guaranteed SAT instances (2000)

12. Phase Transition Almost all unsolvable area Fraction of pre-assignment Fraction of unsolvable cases Almost all solvable area Complexity Graph Phase transition 42%50%20%42%50%20% Underconstrained area Critically constrained area Overconstrained area

13. Easy-Hard-Easy pattern in local search % holes Computational Cost Walksat Order 30, 33, 36 “Over” constrained area Underconstrained area

14. Are we ready to predict run times? Problem: high variance log scale

15. Deep structural features Rectangular PatternAligned PatternBalanced Pattern TractableVery hard Hardness is also controlled by structure of constraints, not just the fraction of holes

16. Random versus balanced Balanced Random

17. Random versus balanced Balanced Random

18. Random vs. balanced (log scale) Balanced Random

19. Morphing balanced and random

20. Considering variance in hole pattern

21. Time on log scale

22. Effect of balance on hardness Balanced patterns yield (on average) problems that are 2 orders of magnitude harder than random patterns Expected run time decreases exponentially with variance in # holes per row or column E(T) = C -k  Same pattern (differ constants) for DPPL! At extreme of high variance (aligned model) can prove no hard problems exist

23. Intuitions In unbalanced problems it is easier to identify most critically constrained variables, and set them correctly Backbone variables

24. Are we done? Unfortunately, not quite. While few unbalanced problems are hard, “easy” balanced problems are not uncommon To do: find additional structural features that signify hardness Introspection Machine learning (later this talk) Ultimate goal: accurate, inexpensive prediction of hardness of real-world problems

25. Case study 2: AutoWalksat Problem Instances Solver runtime Learning / Analysis Predictive Model dynamic features control / policy

26. Walksat Choose a truth assignment randomly While the assignment evaluates to false Choose an unsatisfied clause at random If possible, flip an unconstrained variable in that clause Else with probability P (noise) Flip a variable in the clause randomly Else flip the variable in the clause which causes the smallest number of satisfied clauses to become unsatisfied. Performance of Walksat is highly sensitive to the setting of P

28. Shortest expected run time when P is set to minimize McAllester, Selman and Kautz (1997) The Invariant Ratio Mean of the objective function Std Deviation of the objective function 0 1 2 3 4 5 6 7 + 10%

29. Automatic Noise Setting Probe for the optimal noise level Bracketed Search with Parabolic Interpolation No derivatives required Robust to stochastic variations Efficient

30. Hard random 3-SAT

31. 3-SAT, probes 1, 2

32. 3-SAT, probe 3

33. 3-SAT, probe 4

34. 3-SAT, probe 5

35. 3-SAT, probe 6

36. 3-SAT, probe 7

37. 3-SAT, probe 8

38. 3-SAT, probe 9

39. 3-SAT, probe 10

40. Summary: random, circuit test, graph coloring, planning

41. Other features still lurking clockwise – add 10%counter-clockwise – subtract 10% More complex function of objective function? Mobility? (Schuurmans 2000)

42. Case Study 3: Restart Policies Problem Instances Solver static features runtime Learning / Analysis Predictive Model dynamic features resource allocation / reformulation control / policy

43. Background Backtracking search methods often exhibit a remarkable variability in performance between: different heuristics same heuristic on different instances different runs of randomized heuristics

44. Cost Distributions Observation (Gomes 1997): distributions often have heavy tails infinite variance mean increases without limit probability of long runs decays by power law (Pareto-Levy), rather than exponentially (Normal) Very short Very long

45. Randomized Restarts Solution: randomize the systematic solver Add noise to the heuristic branching (variable choice) function Cutoff and restart search after a some number of steps Provably eliminates heavy tails Very useful in practice Adopted by state-of-the art search engines for SAT, verification, scheduling, …

46. Effect of restarts on expected solution time (log scale)

47. How to determine restart policy Complete knowledge of run-time distribution (only): fixed cutoff policy is optimal (Luby 1993) argmin t E(R t ) where E(R t ) = expected soln time restarting every t steps No knowledge of distribution: O(log t) of optimal using series of cutoffs 1, 1, 2, 1, 1, 2, 4, … Open cases addressed by our research Additional evidence about progress of solver Partial knowledge of run-time distribution

48. Backtracking Problem Solvers Randomized SAT solver Satz-Rand, a randomized version of Satz ( Li & Anbulagan 1997) DPLL with 1-step lookahead Randomization with noise parameter for increasing variable choices Randomized CSP solver Specialized CSP solver for QCP ILOG constraint programming library Variable choice, variant of Brelaz heuristic

49. Formulation of Learning Problem Different formulations of evidential problem Consider a burst of evidence over initial observation horizon Observation horizon + time expended so far General observation policies LongShort Observation horizon Median run time 1000 choice points Observation horizon

50. Observation horizon + Time expended Formulation of Learning Problem Different formulations of evidential problem Consider a burst of evidence over initial observation horizon Observation horizon + time expended so far General observation policies LongShort Observation horizon Median run time 1000 choice points t1t1t1t1 t2t2t2t2 t3t3t3t3

51. Formulation of Dynamic Features No simple measurement found sufficient for predicting time of individual runs Approach: Formulate a large set of base-level and derived features Base features capture progress or lack thereof Derived features capture dynamics 1 st and 2 nd derivatives Min, Max, Final values Use Bayesian modeling tool to select and combine relevant features

52. CSP : 18 basic features, summarized by 135 variables # backtracks depth of search tree avg. domain size of unbound CSP variables variance in distribution of unbound CSP variables Satz : 25 basic features, summarized by 127 variables # unbound variables # variables set positively Size of search tree Effectiveness of unit propagation and lookahead Total # of truth assignments ruled out Degree interaction between binary clauses, Dynamic Features

53. Single instance Solve a specific instance as quickly as possible Learn model from one instance Every instance Solve an instance drawn from a distribution of instances Learn model from ensemble of instances Any instance Solve some instance drawn from a distribution of instances, may give up and try another Learn model from ensemble of instances Different formulations of task

54. Sample Results: CSP-QWH-Single QWH order 34, 380 unassigned Observation horizon without time Training: Solve 4000 times with random Test: Solve 1000 times Learning: Bayesian network model MS Research tool Structure search with Bayesian information criterion (Chickering, et al. ) Model evaluation: Average 81% accurate at classifying run time vs. 50% with just background statistics (range of 98% - 78%)

56. Learned Decision Tree Min of 1 st derivative of variance in number of uncolored cells across columns and rows. Min number of uncolored cells averaged across columns. Min depth of all search leaves of the search tree. Change of sign of the change of avg depth of node in search tree. Max in variance in number of uncolored cells.

57. Restart Policies Model can be used to create policies that are better than any policy that only uses run-time distribution Example: Observe for 1,000 steps If “run time > median” predicted, restart immediately; else run until median reached or solution found; If no solution, restart. E(R fixed ) = 38,000 but E(R predict ) = 27,000 Can sometimes beat fixed even if observation horizon > optimal fixed !

58. Ongoing work Optimal predictive policies Dynamic features + Run time + Static features Partial information about run time distribution E.g.: mixture of two or more subclasses of problems Cheap approximations to optimal policies Myoptic Bayes

59. Conclusions Exciting new direction for improving power of search and reasoning algorithms Many knobs to learn how to twist Noise level, restart policies just a start Lots of opportunities for cross-disciplinary work Theory Machine learning Experimental AI and OR Reasoning under uncertainty Statistical physics