Report time: On July 6, 2016 (Wednesday) 3:30p.m.

Report place: The seventh conference room in Jingshi Academic Hall

Reporter: Qingyang Zhang, University of Arkansas

Abstract: The hidden Markov model (HMM) is a popular tool in modeling sequential events in many applications. A conventional discrete HMM implicitly assumes that each state has a geometric duration distribution, which can be incorrect in real applications. In this talk, we consider a duration HMM (dHMM) where the state duration distribution is unknown but of scientific interest. Most existing methods on duration estimation of dHMM suffers from extreme instability (converging distribution can migrate to wildly biased distribution) and over-specification for the parametric form of duration distribution. To overcome these difficulties, we proposed a robust nonparametric model, where the duration distribution for each state is iteratively estimated using the posterior probability combined with a kernel smoothing method. In addition, we proposed a generic scaling method to avoid underflow issue for forward and backward algorithm where logarithm transformation does not apply. These methods are illustrated with applications in prediction of nucleosome spacing distribution in two eukaryotic genomes.