An Empirical Comparison of CSP Decomposition Methods.pdf

An Empirical Comparison of CSP Decomposition Methods Sathiamoorthy Subbarayan Computational Logic and Algorithms Group IT University of Copenhagen Denmark sathi@itu.dk Abstract. We present an empirical comparison of decomposition tech- niques for CSPs. We compare three popular heuristics for tree decompo- sition and an exact method for optimal tree decomposition, along with two methods for hypertree decomposition. We use a small sample of instances in the experiments. We find that the connected hypertree de- composition method finds more optimal width decompositions than the other methods. In the case, the width of the tree decomposition obtained is the criteria for comparison, then the heuristics are still the better op- tion than the optimal decomposition methods. Especially, the min-fill heuristic quickly finds relatively very good decompositions. 1 Introduction The CSPs from many problem domains have an underlying tree-like structure. The tree-likeness of a CSP can be captured by the treewidth [1] of the constraint graph (or hypergraph) [2] of the CSP. We have observed low treewidth in many instances from several problem domains, like configuration, fault-trees, protein- side chain packing, Bayesian networks and digital circuits. When a CSP instance is known to have a low treewidth, it is often advanta- geous to exploit the low treewidth by obtaining a tree decomposition [1] of the CSP. The obtained tree decomposition can be used to speed-up constraint solv- ing. A well-known technique for exploiting tree decomposition for CSP solving is the tree-clustering [2] method. The speed-up is often high in applications like exact-inference in Bayesian networks and configuration, where there is a need to compile all the solutions. The usual method for obtaining a tree-decomposition is to use one of the several heuristics on the constraint graph of an instance. The popular heuristics are maximum cardinality search (mcs) [3], minimum degree (mindeg) [4] and minimum fill-in