A New Monte Carlo Sampling Method

Speaker:Dr. Ning Jianhui

Time: 2:00 pm - 3:00 pm, Apr. 26(Thursday), 2012

Venue: E301

Abstract:

In many fields of science, we may need to generate the sample of a given density, some times the density may be very complicate or only known up to a multiplicative constant. Many methods are proposed to solve this problem, for example the inverse of cumulative distribution function method, accept-reject method, sampling/importance resampling method, and currently the most popular method: Markov Chain Monte Carlo sampler. In this research (cooperated with Prof. Wang from Rutgers university, Prof. Fang from UIC, and Dr. Zhou from Sichuan University), we propose a new sampler, which can generate samples from most of probability kernel. the samples generated by our sampler exhibit little autocorrelation. And the simulation results also show its accuracy and reproducibility are comparable with the pseudo-random samples. Here, reproducibility is measured via the Monte Carlo error. Comparisons between new sampler and the Metropolis-Hasting algorithm, the Gibbs sampler and the slice sampler will also be shown in my presentation.