Sampling and Soundness: Can We Have Both?

Recent research on model counting in CNF formulas has shown that a certain sampling method can yield results that are sound with a provably high probability. The key idea is to iteratively restrict the search space, and to randomly choose which part to consider. The expected value of this sampled count is equal to the real count. If one minimizes over several trials, and purposefully underestimates the outcome of each trial by a constant factor, then the probability that the sampled count exceeds the real count decreases exponentially in the number of trials. This method has proven to be quite successful for many CNF formulas. The big question is: Can we devise similar methods for reasoning in the Semantic Web? Is it possible to obtain provably high-quality results based on sampling? 1 Sampling and Model Counting Recent research on model counting in propositional CNFs [2, 1] has shown that a certain sampling method can yield results that are sound with a provably high probability. The ...

• Type: Conference
• Year: 2007
• Where: SEMWEB
• Authors: Carla P. Gomes, Jörg Hoffmann, Ashish Sabharwal, Bart Selman
• Keywords: Certain Sampling Method, Cnf Formula, Model Counting, SEMWEB 2007