Performing Incremental Bayesian Inference by Dynamic Model Counting.pdfVIP

Performing Incremental Bayesian Inference by Dynamic Model Counting Wei Li and Peter van Beek and Pascal Poupart School of Computer Science University of Waterloo Waterloo, Ontario N2L 3G1, Canada {w22li, vanbeek, ppoupart}@cs.uwaterloo.ca Abstract In this paper, we propose dynamic model counting (Dy- naMC). The philosophy behind dynamic model counting is The ability to update the structure of a Bayesian net- intuitive. Bayesian networks can be efficiently compiled into work when new data becomes available is crucial for CNFs (Darwiche 2002; Sang, Beame, Kautz 2005). When building adaptive systems. Recent work by Sang, Beame, and Kautz (AAAI 2005) demonstrates that the two CNF formulas share many common clauses, it is possi- well-known Davis-Putnam procedure combined with a ble to use the knowledge learned while counting the solu- dynamic decomposition and caching technique is an ef- tions of the first formula to simplify the solving and count- fective method for exact inference in Bayesian networks ing of the second formula. We show that component caching with high density and width. In this paper, we define (good learning) can significantly improve Bayesian infer- dynamic model counting and extend the dynamic de- ence and other encoded dynamic model counting problems. composition and caching te