## Exploiting the deep structure of constraint problems (1994)

Venue: | Artificial Intelligence |

Citations: | 76 - 8 self |

### BibTeX

@ARTICLE{Williams94exploitingthe,

author = {Colin Williams and Tad Hogg},

title = {Exploiting the deep structure of constraint problems},

journal = {Artificial Intelligence},

year = {1994},

volume = {70},

pages = {73--117}

}

### Years of Citing Articles

### OpenURL

### Abstract

We introduce a technique for analyzing the behavior of sophisticated A.I. search programs working on realistic, large-scale problems. This approach allows us to predict where, in a space of problem instances, the hardest problems are to be found and where the fluctuations in difficulty are greatest. Our key insight is to shift emphasis from modelling sophisticated algorithms directly to modelling a search space that captures their principal effects. We compare our model’s predictions with actual data on real problems obtained independently and show that the agreement is quite good. By systematically relaxing our underlying modelling assumptions we identify their relative contribution to the remaining error and then remedy it. We also discuss further applications of our model and suggest how this type of analysis can be generalized to other kinds of A.I. problems. Chapter 1

### Citations

3882 | Optimization by simulated annealing
- Kirkpatrick
- 1983
(Show Context)
Citation Context ...rch. This raises the question of the applicability of these results to other search methods, in particular those that operate on complete states such as heuristic repair [24, 25], simulated annealing =-=[21]-=- and GSAT [36]. We saw that the cost for backtrack search was determined by the bulge in the number of goods, which occurs well below the solution level in the lattice. By contrast, it is not obvious ... |

2489 |
Equation of state calculations by fast computing machines
- Metropolis, Rosenbluth, et al.
- 1953
(Show Context)
Citation Context ...of size 2 drawn from a distribution in which those Sperner systems with average overlap θ = 1 � m(m − 1) s�=s ′ |s ∩ s ′ | occur with relative probability e αθ using the standard Metropolis algorithm =-=[22]-=-. Positive α values bias the selection toward Sperner systems with higher than random overlap, and conversely negative values bias the selection toward those with lower overlap. For comparison, when α... |

1905 | Random Graphs
- Bollobás
- 2001
(Show Context)
Citation Context ... 2 is equivalent to whether the largest set of nodes with no edges between them in the corresponding graph (the so-called independence number of the graph) has at least µ nodes. While it can be shown =-=[2]-=- that this is almost always true at least up to β = 5.57 and almost never true above βcrit = 9.3 for the example considered here, the question of precisely where the transition point is remains open. ... |

719 |
Heuristics: Intelligent Search Strategies for Computer Problem Solving
- Pearl
- 1985
(Show Context)
Citation Context ...nsion or combination of partial solutions until a complete solution is found, backtracking whenever it is impossible to extend a partial solution further. This includes a variety of search algorithms =-=[32]-=- that range from naive backtrack, with a random ordering for the variables and returning to the most recent decision point during backtrack, to sophisticated heuristic methods that attempt to select g... |

700 | A new method for solving hard satisfiability problems
- Selman, Levesque, et al.
- 1992
(Show Context)
Citation Context ...es the question of the applicability of these results to other search methods, in particular those that operate on complete states such as heuristic repair [24, 25], simulated annealing [21] and GSAT =-=[36]-=-. We saw that the cost for backtrack search was determined by the bulge in the number of goods, which occurs well below the solution level in the lattice. By contrast, it is not obvious how structural... |

604 | Where the Really Hard Problems Are
- Cheeseman, Kanefsky, et al.
- 1991
(Show Context)
Citation Context ...sChapter 1 Introduction In recent years, there has been a wave of experimental results reporting how different constraint satisfaction algorithms typically perform across a range of problem instances =-=[35, 13, 4, 30, 26]-=-. These studies reveal the existence of an easy-hard-easy pattern in computational cost, as a parameter, that distinguishes different classes of problem instances, is varied. The fact that this patter... |

584 |
Generalization as search
- Mitchell
- 1982
(Show Context)
Citation Context ...d search spaces arise. For example, a classic approach to concept learning (the Version Space) exploits upper and lower hulls on a lattice as a compact representation for a space of possible concepts =-=[16, 27]-=-. In this interpretation the lattice nodes are variants of an emerging concept labelled as positive, negative or unseen instances of it. Structural properties of interest are, e.g., the size of the up... |

532 |
Optimization by simulated annealing: An experimental evaluation. Part 1, graph partitioning
- Johnson, Aragon, et al.
- 1989
(Show Context)
Citation Context ...ost constrained unassigned variable had no possible values, then no extensions were possible and the search backtracked. In the case of graph coloring, this method corresponds to the Brelaz heuristic =-=[20]-=- used for the empirical data discussed in §4. The measured search cost is the number of nodes examined. Because this heuristic only instantiates variables with at least one remaining consistent value,... |

465 |
An assumption-based TMS
- Kleer
- 1986
(Show Context)
Citation Context ...cation” as being some mechanism whereby redundant and irrelevant search is avoided. A particular example is the lattice of assumptions that assumption-based truth maintenance systems (ATMSs) navigate =-=[9]-=-. This allows us to finesse handling the minutiae of real algorithms and yet still make quantitatively accurate predictions concerning real problems. The second key ingredient is the observation that ... |

437 | Partial Constraint Satisfaction
- Freuder, Wallace
- 1992
(Show Context)
Citation Context ...sficing searches, the focus is on finding high quality, but not necessarily optimal, states rapidly. In a CSP context, the quality of a state could be characterized by how few constraints it violates =-=[14]-=- (perhaps weighted by a measure of each constraint’s importance). More generally, where one is interested in a “high quality” solution rather than a “correct” solution one could associate a quality va... |

421 | Minimizing conflicts: A heuristic repair method for constraint satisfaction and scheduling problems
- Minton, Johnston, et al.
- 1992
(Show Context)
Citation Context ...the context of a backtrack search. This raises the question of the applicability of these results to other search methods, in particular those that operate on complete states such as heuristic repair =-=[24, 25]-=-, simulated annealing [21] and GSAT [36]. We saw that the cost for backtrack search was determined by the bulge in the number of goods, which occurs well below the solution level in the lattice. By co... |

312 |
Quantitative results concerning the utility of explanation-based learning
- Minton
- 1990
(Show Context)
Citation Context ...for avoiding redundant and irrelevant computation. Redundant computation can be avoided by caching past successes (directly or as generalizations) and reusing them as and when they subsequently arise =-=[12, 23]-=-. Similarly, irrelevant computation can be avoided by caching impossibilities and prohibiting further exploration of inferences dependent upon them [9, 10, 12]. The exact cost to solve a CSP depends o... |

290 |
and easy distributions of SAT problems
- Hard
- 1992
(Show Context)
Citation Context ...sChapter 1 Introduction In recent years, there has been a wave of experimental results reporting how different constraint satisfaction algorithms typically perform across a range of problem instances =-=[35, 13, 4, 30, 26]-=-. These studies reveal the existence of an easy-hard-easy pattern in computational cost, as a parameter, that distinguishes different classes of problem instances, is varied. The fact that this patter... |

280 |
Advanced mathematical methods for scientists and engineers. International series in pure and applied mathematics
- Bender, Orszag
- 1978
(Show Context)
Citation Context ...ue of ˆη decreases, i.e., the maximum number of goods occurs at lower levels in the lattice. Because of the increasing dominance of the largest terms in the sum, it can be approximated by an integral =-=[1]-=- whose leading behavior is simply � � � � �� k ˆη ln 〈C〉 ∼ µ ˆη ln b + β ln 1 − (3.8) b 10 8 beta 6 4 (3.6)sThis decreases as more nogoods are added for two reasons: first, the maximum point moves low... |

202 | Experimental results on the crossover point in satis problems
- Crawford, Auton
- 1993
(Show Context)
Citation Context ...aph coloring and SAT [4, 5]), a number of exact bounds on the location of the transition point (e.g., for the chromatic number of graphs [2]), and exact evaluation for small-size problems such as SAT =-=[8]-=-. Regrettably, with the exception some insightful work by Provan [34], little attention has been paid to unravelling the properties of families of minimized nogoods. Provan’s model differs from ours b... |

178 |
The mathematics of inheritance systems
- Touretzky
- 1986
(Show Context)
Citation Context ...ions to this CSP. We should note that these lattice structures correspond to the sets of assumptions that an ATMS creates [9]. Similar lattice and inheritance structures recur throughout A.I. systems =-=[38]-=- and, although the interpretation of the nodes and links may change, the basic structure remains the same. 2.2 Realistic Lattices Each possible set of minimized nogoods gives rise to a corresponding s... |

156 |
Solving large-scale constraint satisfaction and scheduling problems using a heuristic repair method
- Minton, Johnston, et al.
- 1990
(Show Context)
Citation Context ...the context of a backtrack search. This raises the question of the applicability of these results to other search methods, in particular those that operate on complete states such as heuristic repair =-=[24, 25]-=-, simulated annealing [21] and GSAT [36]. We saw that the cost for backtrack search was determined by the bulge in the number of goods, which occurs well below the solution level in the lattice. By co... |

139 |
Trial by fire: Understanding the design requirements for agents in complex environments
- Cohen, Greenberg, et al.
- 1989
(Show Context)
Citation Context ...reliable estimates of properties of interest can be made. This is fortunate because realistic systems, e.g., those that interact with complex and unpredictable environments such as the physical world =-=[7]-=-, may defy exact specification. In our approach only a few properties of the problem are specified and we assume the others take on values according to some probability distribution, allowing the law ... |

99 |
On selecting a satisfying truth assignment
- Papadimitriou
- 1991
(Show Context)
Citation Context ...sChapter 1 Introduction In recent years, there has been a wave of experimental results reporting how different constraint satisfaction algorithms typically perform across a range of problem instances =-=[35, 13, 4, 30, 26]-=-. These studies reveal the existence of an easy-hard-easy pattern in computational cost, as a parameter, that distinguishes different classes of problem instances, is varied. The fact that this patter... |

81 |
Search rearrangement backtracking and polynomial average time
- Purdom
- 1983
(Show Context)
Citation Context |

70 |
ProbabiIif~ and statistics
- Papoulis
- 1990
(Show Context)
Citation Context ...articular search algorithm. This step can also be used to give an alternate derivation of the maximum in the cost proxy described above [40]. 3 For a nonnegative random variable X, P (X > t) < 〈X〉 /t =-=[31]-=-. 4 This is an application of the Markov inequality to (X − 〈X〉) 2 giving P (|X − 〈X〉 | ≥ t) ≤ var(X)/t 2 . In our case, we need a bound on the probability to be significantly below the average: P (X ... |

58 |
A comparison of ATMS and CSP techniques
- Kleer
- 1989
(Show Context)
Citation Context ...ystem such that no set is a subset of any other [3]. In this sense a CSP, at its deepest level, becomes fundamentally equivalent to its Sperner system. To illustrate how lattice representations arise =-=[11]-=-, recall that a constraint satisfaction problem (CSP) involves a set of µ variables, {v1, v2, . . . , vµ} each having an associated set of domain values, bv1 , bv2 , . . . , bvµ respectively, together... |

45 |
Probabilistic analysis of the Davis Putnam procedure for solving the satisfiability problem
- Franco, Paull
- 1983
(Show Context)
Citation Context |

43 |
Solving the really hard problems with cooperative search
- HOGG, C
- 1993
(Show Context)
Citation Context ... have led to a number of studies of large-scale behavior of various problems based on specifications involving a small number of parameters [4, 19, 17, 37, 39] as well as applications of such results =-=[15, 18]-=-. However, these studies have all focussed on the superficial description of a problem and not on its deep structure. The danger in such an oversight is that superficially different problems might in ... |

41 | Polynomial-time learning with version spaces
- Hirsh
- 1992
(Show Context)
Citation Context ...d search spaces arise. For example, a classic approach to concept learning (the Version Space) exploits upper and lower hulls on a lattice as a compact representation for a space of possible concepts =-=[16, 27]-=-. In this interpretation the lattice nodes are variants of an emerging concept labelled as positive, negative or unseen instances of it. Structural properties of interest are, e.g., the size of the up... |

41 | Using deep structure to locate hard problems
- Williams, Hogg
- 1992
(Show Context)
Citation Context ...m structure (i.e., number of solutions) and independent of any particular search algorithm. This step can also be used to give an alternate derivation of the maximum in the cost proxy described above =-=[40]-=-. 3 For a nonnegative random variable X, P (X > t) < 〈X〉 /t [31]. 4 This is an application of the Markov inequality to (X − 〈X〉) 2 giving P (|X − 〈X〉 | ≥ t) ≤ var(X)/t 2 . In our case, we need a bound... |

33 |
Phase transitions in artificial intelligence systems
- Huberman, Hogg
- 1987
(Show Context)
Citation Context ...ey are not behaving randomly enough 1 . These considerations have led to a number of studies of large-scale behavior of various problems based on specifications involving a small number of parameters =-=[4, 19, 17, 37, 39]-=- as well as applications of such results [15, 18]. However, these studies have all focussed on the superficial description of a problem and not on its deep structure. The danger in such an oversight i... |

30 | and T.Hogg, Cooperative problem Solving
- Clearwater
- 1992
(Show Context)
Citation Context ...ormation to the faster ones. Our results suggest that such cooperative methods can greatly speed up problem solving and that they are most effective for the hardest problems near the phase transition =-=[6, 18]-=-. More generally, one can consider the benefit of additional domain-specific heuristics applied to the search. In particular, there is the question of how heuristics can be expected to interact with t... |

30 |
Learning while searching in constraintsatisfaction-problems
- Dechter
- 1986
(Show Context)
Citation Context ...ng them as and when they subsequently arise [12, 23]. Similarly, irrelevant computation can be avoided by caching impossibilities and prohibiting further exploration of inferences dependent upon them =-=[9, 10, 12]-=-. The exact cost to solve a CSP depends on the search algorithm (e.g., simple backtracking) as well as what is required (e.g., any or all solutions). However, such information is too specific to give ... |

25 |
Phase transitions in sequence matches and nucleic acid structure
- Waterman, Gordon, et al.
- 1987
(Show Context)
Citation Context ...ey are not behaving randomly enough 1 . These considerations have led to a number of studies of large-scale behavior of various problems based on specifications involving a small number of parameters =-=[4, 19, 17, 37, 39]-=- as well as applications of such results [15, 18]. However, these studies have all focussed on the superficial description of a problem and not on its deep structure. The danger in such an oversight i... |

20 |
Observation of phase transitions in spreading activation networks
- Shrager, Hogg, et al.
- 1987
(Show Context)
Citation Context ...ey are not behaving randomly enough 1 . These considerations have led to a number of studies of large-scale behavior of various problems based on specifications involving a small number of parameters =-=[4, 19, 17, 37, 39]-=- as well as applications of such results [15, 18]. However, these studies have all focussed on the superficial description of a problem and not on its deep structure. The danger in such an oversight i... |

14 |
Conspiracy Numbers and Caching for Searching And/Or Trees and Theorem-Proving
- Elkan
- 1989
(Show Context)
Citation Context ...for avoiding redundant and irrelevant computation. Redundant computation can be avoided by caching past successes (directly or as generalizations) and reusing them as and when they subsequently arise =-=[12, 23]-=-. Similarly, irrelevant computation can be avoided by caching impossibilities and prohibiting further exploration of inferences dependent upon them [9, 10, 12]. The exact cost to solve a CSP depends o... |

12 |
Computational Complexity and Phase Transitions
- Cheeseman, Kanefsky, et al.
- 1992
(Show Context)
Citation Context ...asymptotic behavior is zero, i.e., at b ln b βcrit = − (5.12) ln(1 − 1/b) This result can also be obtained more directly by assuming conditional independence among the nogoods introduced by each edge =-=[5]-=-. For the cases of 3 and 4-coloring, this gives βcrit = 8.1 and 19.3, respectively, close to the empirical values given in Table 4.1. 5.2.3 Modelling Simplification: Ensemble with Specified Solution I... |

12 |
On the density of solutions in equilibrium points for the queens problem
- Morris
- 1992
(Show Context)
Citation Context ..., k = 2 (dashed curve, with maximum at β = 9.5) and its mean-field approximation (grey, with maximum at 12). This proxy has been examined, for example, in the specific context of the N-Queens problem =-=[28]-=-. The average number of minimal points can be determined by counting how many ways there are of choosing the minimized nogoods such that the neighbors of a given state at the solution level have at le... |

12 |
How long will it take
- Musick, Russell
- 1992
(Show Context)
Citation Context ...ined, for example, in the context of heuristic repair [24]. This can be elaborated by considering the search as a random walk with various probabilities to reduce the number of conflicts at each step =-=[29]-=-. 31 betas6.3 Additional Applications In this paper we have used the lattice structure of CSPs to predict global properties such as the number of solutions and the search cost. As we outline in this s... |

8 |
Efficiency Analysis of MultipleContext TMSs
- Provan
(Show Context)
Citation Context ...e at an intermediate level in the lattice. This bulge in the expected number of goods as a function of level in the lattice becomes increasingly sharp as µ increases and has been observed empirically =-=[33]-=-. We quantify these behaviors for large problems, i.e., as µ → ∞. The discussion of §2.2 suggests an interesting scaling limit is to keep k fixed and take m = βµ, as summarized in Table 3.2. For behav... |

4 |
A cluster detection algorithm based on percolation theory
- Gotsman
- 1991
(Show Context)
Citation Context ... have led to a number of studies of large-scale behavior of various problems based on specifications involving a small number of parameters [4, 19, 17, 37, 39] as well as applications of such results =-=[15, 18]-=-. However, these studies have all focussed on the superficial description of a problem and not on its deep structure. The danger in such an oversight is that superficially different problems might in ... |

4 |
Phase transitions in highdimensional pattern classification
- Hogg, Kephart
- 1990
(Show Context)
Citation Context |

3 |
The typicality of phase transitions in search
- Williams, Hogg
- 1993
(Show Context)
Citation Context ...ysis of the deep structure of constraint satisfaction problems (CSPs), in terms of 1 Often some correlations in the unspecified degrees of freedom can be tolerated without major effect on the results =-=[41]-=-. 1stheir underlying Sperner systems, and examine its consequences. These include specific phase transitions, in which global properties such as problem difficulty, change abruptly as simple measures ... |