Leveraging- Belief- Propagation,- Backtrack- Search,-and-课件.pptVIP

* * * * * * * * Bounding E[#FMiniCount] Will use confidence interval for E[#FMiniCount] Need samples of the random variable (OK) Need to know the shape of the distribution (?) Key empirical observation: For many problems, the distribution of #FMiniCount is approximately log-normal Can be checked using normality tests (like Shapiro-Wilk) Can be assessed graphically (QQ-plots) In either case, log-normality cannot be guaranteed, only “not disproved” * Generate independent samples from #FMiniCount by independent runs of MiniCount #FMiniCount freq Assess whether the samples come from a log-normal distribution, or a sufficiently similar one Calculate the (one-sided) confidence interval for E[#FMiniCount] and use it as the upper bound for #F Upper bound Mean Sample average might easily underestimate the mean, due to the heavy tail of the distribution MiniCount in a Nutshell * MiniCount Results Very fast! Applicable for mostproblems Sometimes notwell applicable Simple averageoften good, butnot a bound * What Happens When Not Log-Normal? Luckily, often easy to detectin practice Langford problems: * Summary Fast lower bounds with probabilistic guarantees Using an earlier randomized framework Adding the power of Belief Propagation Key improvement: speed! Fast upper bounds with statistical guarantees Upper bounds turn out to be more difficult than lower bounds First efficient way to tackle the problem, albeit not with any probabilistic guarantees * * * * * * * * * * * * * * * * * * * * * * * * * * * * Ashish Sabharwal May 23, 2008 Ashish Sabharwal May 23, 2008 Leveraging Belief Propagation, Backtrack Search, and Statistics for Model Counting Lukas Kroc, Ashish Sabharwal, Bart Selman Cornell University May 23, 2008 CPAIOR-08 Conference, Paris * Talk Outline Model counting / solution counting Problem, motivation, background Lower bound estimates Fast, using belief propagation for guidance Probabilistic correctness guarantees Experiments Upper b