A Generalized Cyclic-Clustering Approach for Solving Structured CSPs

We propose a new method for solving structured CSPs which generalizes and improves the Cyclic-Clustering approach. First, the cutset and the tree-decomposition of the constraint network, which are used for taking advantage of the CSP structure, are computed independently of the notion of triangulated induced subgraph. Then, unlike Cyclic-Clustering, our method can try to solve the tree-decomposition part of the problem without having assigned all the variables of the cutset. Regarding the solving of the tree-decomposition part, we use the BTD method like in. As BTD records and exploits structural (no)goods, we provide some conditions which make possible the use of structural (no)goods recorded during previous calls of BTD and we implement them in a dedicated version of BTD. By so doing, from a theoretical viewpoint, we can provide a theoretical time complexity bound related to parameters of the cutset and the tree-decomposition and, from a practical viewpoint we expect to detect failures earlier and to avoid more redundancies in the search.