Algorithms for the Minimal Network of a CSP and a Classifier for Choosing Between Them

The minimal constraint network of a constraint satisfaction problem (CSP) is a compiled version of the problem where every tuple in a constraint’s relation appears in at least one solution to the CSP. Recently, Gottlob argued that, when a CSP has this property, a number of NP-hard queries can be answered in polynomial time, but he also showed that deciding whether or not a given network is minimal is NP-complete [Gottlob, 2011]. We propose two search-based algorithms for computing the minimal network of a CSP. We investigate the performance of the two algorithms and propose a classifier to select the appropriate algorithm that minimizes the CPU time, using a number of parameters. Our approach constitutes a significant contribution towards the automation of the selection of the appropriate algorithms for computing the minimal network of a CSP. In most cases that we studied, we achieved classifier accuracy of above 90%, which allowed us realize significant time savings