Monte Carlo algorithms on distributed architectures
The collection of vast amounts of data and cheap access to multiple computer cores are important aspects of the modern scientific era. This leads to the necessity of developing distributed algorithms to tackle such problems. Markov chain Monte Carlo (MCMC) algorithms are a very popular class of algorithms which are used in Bayesian statistics and elsewhere. However, these lose attractive theoretical guarantees when implemented on parallel architectures. Sequential Monte Carlo (SMC) algorithms are an alternate class which preserves theoretical guarantees on distributed architectures. I will review some aspects of scaling up MCMC algorithms to big data and introduce distributed SMC algorithms. I will also talk about the optimal way of distributing computing resources across a network in this context.
Refreshments at 3:45pm in Snedecor 2101.