Title: An invitation to adaptive MCMC convergence theory
Abstract: Adaptive Markov chain Monte Carlo (MCMC) algorithms, which automatically tune their parameters based on past samples, have proved extremely useful in practice. There is a lot of freedom how adaptive MCMC samplers can be designed, but some care is needed to ensure that they remain valid (i.e. the Monte Carlo averages are consistent). The self-tuning mechanism makes the algorithms 'non-Markovian', which means that the validity cannot be ensured by standard Markov chains theory. We discuss a simple proof which ensures the consistency of averages, based on a martingale decomposition due to Andrieu and Moulines (Ann. Appl. Probab. 2006). We focus on uniformly ergodic Markov chains and aim to keep the discussion accessible for audience having limited prior knowledge of Markov chains.
The talk is based on a joint work with Pietari Laitinen
Observe that the starting time is 9.00am sharp.
