-
Notifications
You must be signed in to change notification settings - Fork 0
Anthony Lee
One challenge in inference for large, or otherwise complex, datasets is the design of algorithms that scale on parallel and distributed architectures. In the context of Sequential Monte Carlo (SMC) for hidden Markov models, longer data records necessitate an increased number of particles, N, in order to provide accurate estimates: when N is large one is naturally drawn to distributed implementations of the algorithm with particles simulated on different computers on a network. I will discuss alpha SMC, a generalization of Sequential Monte Carlo in which interaction between particles is modulated by a sequence of "alpha" matrices. Theoretical results then motivate the design of adaptive choices of alpha matrices that satisfy a generalized effective sample size criterion. Combined with a useful lower bound on this effective sample size that does not require the transmission of all particle weights to a central processor, this motivates forest resampling - a specific implementation of alpha SMC amenable to implementation in a distributed environment.
Papers:
http://www.e-publications.org/ims/submission/BEJ/user/submissionFile/16911?confirm=551ef11b
http://www2.warwick.ac.uk/fac/sci/statistics/staff/academic-research/lee/sadm_preprint.pdf