This is an announcement of the Statistics departmental seminar. Coffee and refreshment will be served at 3:15pm.

Speaker: Aaron Smith, Tutte Institute for Mathematics and Computing

Time: 3:15-4:30pm on Thursday, Jan 30, 2014.
Location: Sidney Smith Hall, Room 1074

Title: Efficiency Bounds and Concentration Inequalities for Adaptive Samplers

Markov chain Monte Carlo (MCMC) is a ubiquitous tool for estimating integrals over complicated probability distributions. In practice, the performance of MCMC algorithms depends heavily on a large number of tuning parameters that can be difficult to select. This problem is sometimes solved by using “adaptive” MCMC methods to learn parameters on the fly. Although these methods are popular, very little is known about the properties of estimates that they produce. In this talk, I present new finite-time error bounds and concentration inequalities for a popular class of adaptive algorithms, the equi-energy (EE) sampler. These ideas are also used to provide the first proofs that the EE sampler can be more efficient than its non-adaptive competitors.