Efficient Recognition of Acyclic Clustered Constraint Satisfaction Problems

In this paper we present a novel approach to solving Constraint Satisfaction Problems whose constraint graphs are highly clustered and the graph of clusters is close to being acyclic. Such graphs are encountered in many real world application domains such as configuration, diagnosis, model-based reasoning and scheduling. We present a class of variable ordering heuristics that exploit the clustered structure of the constraint network to inform search. We show how these heuristics can be used in conjunction with nogood learning to develop efficient solvers that can exploit propagation based on either forward checking or maintaining arc-consistency algorithms. Experimental results show that maintaining arc-consistency alone is not competitive with our approach, even if nogood learning and a well known variable ordering are incorporated. It is only by using our cluster-based heuristics can large problems be solved efficiently. The poor performance of maintaining arc-consistency is somewhat surprising, but quite easy