A polynomial time computation of the exact correlation structure of k-satisfiability landscapes

The autocorrelation function and related correlation length are statistical quantities that capture the ruggedness of the fitness landscape: a measure that is directly related to the hardness of a problem for certain heuristic search algorithms. Typically, these quantities are estimated empirically by sampling along a random walk. In this paper, we show that a polynomial-time Walsh decomposition of the k-satisfiability evaluation function allows us to compute the exact autocorrelation function and correlation length for any given k-satisfiability instance. We also use the decomposition to compute a theoretical expectation for the autocorrelation function and correlation length over the ensemble of instances generated uniformly at random. We find that this expectation is invariant to the constrainedness of the problem as measured by the ratio of clauses to variables. However, we show that filtered problems, which are typically used in local search studies, have a bias that causes a significant deviation from the expected correlation structure of unfiltered, uniformly generated problems.

[1]  Peter F. Stadler,et al.  Amplitude Spectra of Fitness Landscapes , 1998, Adv. Complex Syst..

[2]  E. Weinberger,et al.  Correlated and uncorrelated fitness landscapes and how to tell the difference , 1990, Biological Cybernetics.

[3]  Alan Smaill,et al.  Backbone Fragility and the Local Search Cost Peak , 2000, J. Artif. Intell. Res..

[4]  Bernd Freisleben,et al.  Fitness landscape analysis and memetic algorithms for the quadratic assignment problem , 2000, IEEE Trans. Evol. Comput..

[5]  Kevin Leyton-Brown,et al.  Hierarchical Hardness Models for SAT , 2007, CP.

[6]  A. Hartmann.  Phase Transitions in Combinatorial Optimization Problems - Basics, Algorithms and Statistical Mechanics , 2005 .

[7]  Alan Smaill,et al.  Backbone FragilityCauses the Local Search Cost , 1999 .

[8]  P. Stadler,et al.  Random field models for fitness landscapes , 1999 .

[9]  Vassilis Zissimopoulos,et al.  On the Classification of NP-complete Problems in Terms of Their Correlation Coefficient , 2000, Discret. Appl. Math..

[10]  Peter C. Cheeseman,et al.  Where the Really Hard Problems Are , 1991, IJCAI.

[11]  Makoto Yokoo.  Why Adding More Constraints Makes a Problem Easier for Hill-climbing Algorithms: Analyzing Landscapes of CSPs , 1997, CP.

[12]  Hector J. Levesque,et al.  Hard and Easy Distributions of SAT Problems , 1992, AAAI.

[13]  P. Stadler.  Landscapes and their correlation functions , 1996 .

[14]  L. Darrell Whitley,et al.  A Tractable Walsh Analysis of SAT and its Implications for Genetic Algorithms , 1998, AAAI/IAAI.

[15]  Thomas Stützle,et al.  Search Space Features Underlying the Performance of Stochastic Local Search Algorithms for MAX-SAT , 2004, PPSN.

[16]  P. Stadler,et al.  The landscape of the traveling salesman problem , 1992 .

[17]  M. Eigen,et al.  Molecular quasi-species. , 1988 .

[18]  Toby Walsh,et al.  Local Search and the Number of Solutions , 1996, CP.

[19]  Andrew J. Parkes,et al.  Clustering at the Phase Transition , 1997, AAAI/IAAI.

[20]  Tad Hogg,et al.  Exploiting the Deep Structure of Constraint Problems , 1994, Artif. Intell..

[21]  S Kirkpatrick,et al.  Critical Behavior in the Satisfiability of Random Boolean Expressions , 1994, Science.