Repulsive Parallel MCMC for Motif Discovery

• Activity Number: 248
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Hisaki Ikebata*+ and Ryo Yoshida
• Companies: The Graduate University for Advanced Studies and The Institute of Statistical Mathematics
• Keywords: MCMC ; Bayesian Statistics ; Bioinformatics ; Motif Discovery

Abstract

A conventional MCMC sampler tends to get stuck in a local mode of a multimodal distribution, and encounters difficulty in moving to different modes within finite time of simulation run. To reach to as many as existing modes, the sampling is often repeated several times with different initial states. Even in doing so, however, a problem would remain; the different chains get stuck to one particular mode having a much higher probability mass. To overcome this difficulty, we propose a new parallel MCMC algorithm. This idea is rather simple: During a parallel run of several MCMC simulations, a repulsive effect is added on each pair of the trajectories. Then these samplers are forced to apart from other samplers by the repulsive effect and can explore different regions, i.e. a task allocation. With just one-time parallel computation, we would be able to capture as diverse as many modes. After describing the methodology, we will demonstrate an application to a pattern mining problem for DNA sequences called "motif finding problem".