## Using CSP Look-Back Techniques to Solve Exceptionally Hard SAT Instances (1996)

### Cached

### Download Links

- [www.bayardo.org]
- [www.almaden.ibm.com]
- [www.almaden.ibm.com]
- DBLP

### Other Repositories/Bibliography

Venue: | Principles and Practice of Constraint Programming |

Citations: | 36 - 2 self |

### BibTeX

@INPROCEEDINGS{Bayardo96usingcsp,

author = {Roberto J. Bayardo and Robert Schrag},

title = {Using CSP Look-Back Techniques to Solve Exceptionally Hard SAT Instances},

booktitle = {Principles and Practice of Constraint Programming},

year = {1996},

pages = {46--60},

publisher = {Springer}

}

### Years of Citing Articles

### OpenURL

### Abstract

Abstract. While CNF propositional satisfiability (SAT) is a sub-class of the more general constraint satisfaction problem (CSP), conventional wisdom has it that some well-known CSP look-back techniques-- including backjumping and learning-- are of little use for SAT. We enhance the Tableau SAT algorithm of Crawford and Auton with look-back techniques and evaluate its performance on problems specifically designed to challenge it. The Random 3-SAT problem space has commonly been used to benchmark SAT algorithms because consistently difficult instances can be found near a region known as the phase transition. We modify Random 3-SAT in two ways which make instances even harder. First, we evaluate problems with structural regularities and find that CSP look-back techniques offer little advantage. Second, we evaluate problems in which a hard unsatisfiable instance of medium size is embedded in a larger instance, and we find the look-back enhancements to be indispensable. Without them, most instances are “exceptionally hard ”-orders of magnitude harder than typical Random 3-SAT instances with the same surface characteristics.

### Citations

1152 |
A Computing Procedure for Quantification Theory
- Davis, Putnam
- 1960
(Show Context)
Citation Context ...s over the values { true, false} . For SAT, the most popular systematicsalgorithms are variants of the Davis-Logemann-Loveland modification [8] to the procedure originally defined by Davis and Putnam =-=[7]-=-; hereafter we refer to this procedure as “DP”. In CSP terms, the procedure is equivalent to backtrack search with forward checking and an ordering heuristic favoring unit-domained variables. Two effe... |

790 |
D.: A machine program for theorem proving
- Davis, Logemann, et al.
- 1962
(Show Context)
Citation Context ...SAT) is a specific kind of CSP in which every variable ranges over the values { true, false} . For SAT, the most popular systematicsalgorithms are variants of the Davis-Logemann-Loveland modification =-=[8]-=- to the procedure originally defined by Davis and Putnam [7]; hereafter we refer to this procedure as “DP”. In CSP terms, the procedure is equivalent to backtrack search with forward checking and an o... |

713 | A New Method for Solving Hard Satisfiability Problems
- Selman, Levesque, et al.
- 1992
(Show Context)
Citation Context ...sts these instances may have properties similar to (though not as pronounced as) those from Exceptionally-Hard 3-SAT. Mazure et al. [22] recently developed a look-ahead technique for DP based on GSAT =-=[26]-=-. The technique selects branch variables by counting the number of times a variable occurs in clauses falsified by assignments made during a GSAT search on the current sub-problem. The intent is to fo... |

377 | Dynamic backtracking - Ginsberg - 1993 |

375 | Hybrid Algorithms For The Constraint Satisfaction Problem
- Prosser
- 1993
(Show Context)
Citation Context ...raint recording). In CSP algorithms, techniques from both classes are popular; for instance, one common combination of techniques (e.g. [1, 12, 28]) is forward checking, conflict-directed backjumping =-=[23]-=-, and an ordering heuristic preferring variables with the smallest domains. CNF propositional satisfiability (SAT) is a specific kind of CSP in which every variable ranges over the values { true, fals... |

285 |
Enhancement schemes for constraint processing: backjumping, learning, and cutset decomposition
- Dechter
- 1990
(Show Context)
Citation Context ...ceslimits on derived clauses [2,16], or some method for efficiently producing smaller culprit sets than conflict-directed backjumping (e.g. possibly along the lines of Dechter’s deep learning schemes =-=[9]-=-). 6 Related and Future Work Ginsberg & McAllester [16] evaluate a CSP algorithm they call “partial-order dynamic backtracking” on a 3-SAT problem space with restricted structure. This problem space c... |

262 |
Forward Reasoning and Dependency-Directed Backtracking in a System for Computer-Aided Circuit Analysis
- Stallman, Sussman
- 1977
(Show Context)
Citation Context ...applying conflict-directed backjumping (CBJ) [23], another CBJ with third-order learning [12], and the last CBJ with unrestricted learning (sometimes referred to as “dependency-directed backtracking” =-=[30]-=-). As look-back techniques, backjumping and learning are invoked when the algorithm reaches a failure point where at least one variable assignment must be undone before search can progress. Both explo... |

207 | Experimental results on the cross-over point in satis problems
- Crawford, Auton
- 1993
(Show Context)
Citation Context ...tage points. Tableau, while performing worse on these instances than typical Random 3-SAT instances, has a respectable growth rate when compared to naive DP variants. For instance, Crawford and Auton =-=[4]-=- find that DP using a most-constrained-first variablesClauses Branch Points at Crossover Point 800 750 700 650 600 550 500 450 400 Variable-Regular 3-SAT Literal-Regular 3-SAT 350 100 12000 110 120 13... |

171 | Improvements to Propositional Satisfiability Search Algorithms
- Freeman
- 1995
(Show Context)
Citation Context ...procedure is equivalent to backtrack search with forward checking and an ordering heuristic favoring unit-domained variables. Two effective modern variants of this algorithm are Tableau [5] and POSIT =-=[11]-=-, both amounting to DP with highly optimized variable ordering heuristics. Are these SAT algorithms missing anything by not incorporating conflict-directed backjumping or another look-back technique? ... |

155 | Critical behavior in the satisfiability of random Boolean expressions
- Kirkpatrick, Selman
- 1994
(Show Context)
Citation Context ... DP implementation. Baker [1], working with graph coloring CSPs, finds no problems to be extremely hard for a conflict-directed backjumping and unrestricted learning algorithm. Selman and Kirkpatrick =-=[25]-=-, using an earlier version of Tableau [4] and Random 3-SAT, investigate the incidence of EHIs when a given instance is subject to an equivalence-preserving random renaming of its variables. They repor... |

144 | Experimental results on the crossover point in random 3-SAT
- Crawford, Auton
- 1996
(Show Context)
Citation Context ...SP terms, the procedure is equivalent to backtrack search with forward checking and an ordering heuristic favoring unit-domained variables. Two effective modern variants of this algorithm are Tableau =-=[5]-=- and POSIT [11], both amounting to DP with highly optimized variable ordering heuristics. Are these SAT algorithms missing anything by not incorporating conflict-directed backjumping or another look-b... |

123 | Locating the phase transition in binary constraint satisfaction problems
- Smith, Dyer
- 1996
(Show Context)
Citation Context ...e search space to realize inevitable dead-ends as quickly as possible. We were interested in identifying the effects of highly regular instances on Tableau and its enhancements. Various other studies =-=[6, 13, 29, 31]-=- have investigated the effects of increased regularity on SAT and CSP solving, finding that higher regularity increases mean difficulty. While look-back techniques do not significantly improve Tableau... |

119 | A Theoretical Evaluation of Selected Backtracking Algorithms
- Kondrak, Beek
- 1997
(Show Context)
Citation Context ...e culprit identification scheme used by Prosser’s conflict directed backjumping is widely used [1, 12, 28], requires little overhead [28], and is provably more effective than some of its predecessors =-=[20]-=-. Given a culprit identification scheme, the next issue to be decided is how to exploit the culprits. Pure CBJ simply backs up to the most recent culprit to have been assigned a value without recordin... |

104 | Generating Hard Satisfiability Problems
- Selman, Mitchell, et al.
- 1996
(Show Context)
Citation Context ...3-SAT is already a source of consistently hard problem instances -- those in the region of the phase transition occurring when the ratio of constraints to variables increases through a critical value =-=[27]-=-. The phase transition separates an under-constrained region, where almost all instances are satisfiable and easy, from an over-constrained region, where almost all instances are unsatisfiable and rel... |

94 | The Hardest Constraint Problems: A Double Phase Transition
- Hogg, Williams
- 1993
(Show Context)
Citation Context ... regular problems, we find that look-back techniques offer little or no advantage; for solving our embedded problems, we find them indispensable. Researchers working with random spaces for other CSPs =-=[1, 17]-=-, with other SAT problem spaces [14, 15], or with Random 3-SAT but an algorithm other than Tableau [14, 15], have found rare instances in the under-constrained region so difficult as to render the mea... |

83 | Easy problems are sometimes hard
- Gent, Walsh
- 1994
(Show Context)
Citation Context ...ck techniques offer little or no advantage; for solving our embedded problems, we find them indispensable. Researchers working with random spaces for other CSPs [1, 17], with other SAT problem spaces =-=[14, 15]-=-, or with Random 3-SAT but an algorithm other than Tableau [14, 15], have found rare instances in the under-constrained region so difficult as to render the mean difficulty higher there than in the tr... |

80 | D.: A complexity analysis of space-bound learning algorithms for the constraint satisfaction problem
- Bayardo, Miranker
- 1996
(Show Context)
Citation Context ...ber of constraints. We believe that any SAT algorithm applying learning on instances like ours will require either limited order learning as employed here, time or relevanceslimits on derived clauses =-=[2,16]-=-, or some method for efficiently producing smaller culprit sets than conflict-directed backjumping (e.g. possibly along the lines of Dechter’s deep learning schemes [9]). 6 Related and Future Work Gin... |

80 | Dead-end driven learning
- Frost, Dechter
- 1994
(Show Context)
Citation Context ...ligent backtracking) and learning (also known as nogood or constraint recording). In CSP algorithms, techniques from both classes are popular; for instance, one common combination of techniques (e.g. =-=[1, 12, 28]-=-) is forward checking, conflict-directed backjumping [23], and an ordering heuristic preferring variables with the smallest domains. CNF propositional satisfiability (SAT) is a specific kind of CSP in... |

55 | Intelligent backtracking on constraint satisfaction problems: Experimental and theoretical results
- Baker
- 1995
(Show Context)
Citation Context ...ligent backtracking) and learning (also known as nogood or constraint recording). In CSP algorithms, techniques from both classes are popular; for instance, one common combination of techniques (e.g. =-=[1, 12, 28]-=-) is forward checking, conflict-directed backjumping [23], and an ordering heuristic preferring variables with the smallest domains. CNF propositional satisfiability (SAT) is a specific kind of CSP in... |

50 |
Eliminating duplication with the hyper-linking strategy
- Lee, Plaisted
- 1992
(Show Context)
Citation Context ...ring 27 branch points. Some of these instances were difficult for Tableau without learning but with CBJ. We have not yet explored the effect of their technique on our problem spaces. Lee and Plaisted =-=[21]-=- enhance DP with a backjumping scheme similar to (but not as powerful as) CBJ which they use as a subroutine in a first-order theorem proving system. Gent and Walsh [15] experiment with an implementat... |

44 | Sparse constraint graphs and exceptionally hard problems
- Smith, Grant
- 1995
(Show Context)
Citation Context ...ligent backtracking) and learning (also known as nogood or constraint recording). In CSP algorithms, techniques from both classes are popular; for instance, one common combination of techniques (e.g. =-=[1, 12, 28]-=-) is forward checking, conflict-directed backjumping [23], and an ordering heuristic preferring variables with the smallest domains. CNF propositional satisfiability (SAT) is a specific kind of CSP in... |

35 | The Satisfiability Constraint Gap
- Gent, Walsh
- 1996
(Show Context)
Citation Context ...ck techniques offer little or no advantage; for solving our embedded problems, we find them indispensable. Researchers working with random spaces for other CSPs [1, 17], with other SAT problem spaces =-=[14, 15]-=-, or with Random 3-SAT but an algorithm other than Tableau [14, 15], have found rare instances in the under-constrained region so difficult as to render the mean difficulty higher there than in the tr... |

24 | Graph coloring
- Culberson
- 2004
(Show Context)
Citation Context ...find a left-shifted phase transition and increased difficulty. Dubois and Boufkhad also investigate literal-regular instances they call “doubly balanced” in a forthcoming paper [10]. Culberson et al. =-=[6]-=- and Vlasie [31] describe generators for graph coloring CSPs in which graphs are endowed with a variety of structural properties intended to make coloring difficult. Their empirical results cannot be ... |

9 |
Resolution search
- Chvátal
- 1997
(Show Context)
Citation Context ...they call “Constant Probability” and find that the backjumping scheme reduces the incidence of EHIs in the under-constrained region, but fails to eliminate them at large enough problem sizes. Chvatal =-=[3]-=- also applies look-back to SAT through “resolution search” -- a DP variant with what appears to be a novel learning scheme. 1 These instances are available through anonymous FTP at ftp:dimacs.rutgers.... |

9 |
Detecting logical inconsistencies
- Mazure, Sais, et al.
- 1996
(Show Context)
Citation Context ... of the first author: http://www.cs.utexas.edu/users/bayardo/.sjumping and learning enhancements. Related work has identified other 3-SAT instances that are difficult for Tableau or other DP variants =-=[16, 22]-=-. The instances among these which we have tested are trivial for Tableau enhanced with CSP lookback techniques (see Section 6). 2 Definitions SAT involves determining whether a given Boolean expressio... |

6 | Systematic generation of very hard cases for graph 3-colorability
- Vlasie
- 1995
(Show Context)
Citation Context ...e search space to realize inevitable dead-ends as quickly as possible. We were interested in identifying the effects of highly regular instances on Tableau and its enhancements. Various other studies =-=[6, 13, 29, 31]-=- have investigated the effects of increased regularity on SAT and CSP solving, finding that higher regularity increases mean difficulty. While look-back techniques do not significantly improve Tableau... |

3 |
Random Generation of Satisfiable and Unsatisfiable CNF Predicates
- Iwama, Abeta, et al.
- 1992
(Show Context)
Citation Context ...arks/cnf. 2 This is the “intelligent backtracking” algorithm whose implementation they credit to Mark Stickel.sOthers have developed procedures intended to generate hard random problems. Iwama et al. =-=[19]-=- and Rauzy [24] independently generate 3-SAT instances which are known in advance to be satisfiable or unsatisfiable. The authors do not directly compare the difficulty of these problems to problems f... |

2 | Some ideas on random generation of k-SAT instances
- Genisson, Sais
- 1994
(Show Context)
Citation Context ...te 3-SAT instances which are known in advance to be satisfiable or unsatisfiable. The authors do not directly compare the difficulty of these problems to problems from Random 3-SAT. Genisson and Sais =-=[13]-=- investigate 3-SAT generators with controlled distributions of literals -- much like our Literal-Regular 3-SAT -- and also find a left-shifted phase transition and increased difficulty. Dubois and Bou... |

2 | On the Random Generation of 3-SAT Instances
- Rauzy
- 1995
(Show Context)
Citation Context ...s is the “intelligent backtracking” algorithm whose implementation they credit to Mark Stickel.sOthers have developed procedures intended to generate hard random problems. Iwama et al. [19] and Rauzy =-=[24]-=- independently generate 3-SAT instances which are known in advance to be satisfiable or unsatisfiable. The authors do not directly compare the difficulty of these problems to problems from Random 3-SA... |

1 |
From very hard doubly balanced SAT formulae to easy unbalanced SAT formulae, variations of the satisfiability threshold
- Dubois, Boufkhad
- 1996
(Show Context)
Citation Context ... next generation procedure removes this source of irregularity nearly completely. This problem space is equivalent to the “doubly balanced” SAT space investigated independently by Dubois and Boufkhad =-=[10]-=-, and similar to those defined by Genisson and Sais [13]. LITERAL-REGULAR 3-SAT: Inputs are the number of variables ( n ) and the number of clauses ( m ). There are 2n possible literals given n variab... |