Blocksolve a Bottom-Up Approach for Solving Quantified CSPs.pdf

BlockSolve: a Bottom-Up Approach for Solving Quantified CSPs Guillaume Verger and Christian Bessiere LIRMM, CNRS/University of Montpellier, France {verger,bessiere}@lirmm.fr Abstract. Thanks to its extended expressiveness, the quantified con- straint satisfaction problem (QCSP) can be used to model problems that are difficult to express in the standard CSP formalism. This is only recently that the constraint community got interested in QCSP and proposed algorithms to solve it. In this paper we propose BlockSolve, an algorithm for solving QCSPs that factorizes computations made in branches of the search tree. Instead of following the order of the variables in the quantification sequence, our technique searches for combinations of values for existential variables at the bottom of the tree that will work for (several) values of universal variables earlier in the sequence. An ex- perimental study shows the good performance of BlockSolve compared to a state of the art QCSP solver. 1 Introduction The quantified constraint satisfaction problem (QCSP) is an extension of the constraint satisfaction problem (CSP) in which variables are totally ordered and quantified either existentially or universally. This generalization provides a better expressiveness for modelling problems. Model Checking and planning under uncertainty are examples of problems that can nicely be modeled with QCSP. But such an expressiveness has a cost. Whereas CSP is in NP, QCSP is PSPACE-complete. The SAT community has also done a similar generalization from the prob- lem of satisfying a Boolean formula into the quantified Boolean formula problem (QBF). The most natural way to solve instances of QBF or QCSP is to instan- tiate variables from the outermost quantifier to the innermost. This approach is called top-down. Most QBF solvers implement top-down techniques. Those solvers lift SAT techniques to QBF. Nevertheless, Biere [1], or Pan and Vardi [2] proposed different techniques to solve QBF instances.