Adaptive Markov Chain Monte Carlo MSc dissertation @ University of Warwick
Several adaptive MCMC algorithms and their specific features and purposes are introduced. Appropriate performance measures as well as challenging target distributions are presented. Then, a few experiments are designed to test the performance in specific settings. Forcing the acceptance rate to a certain value may result in a reducible chain. Forgetting initial observations from the burn-in phase leads to faster adaptation of the covariance structure. Using heuristic values for some algorithm parameters can strongly affect posterior estimates. Analysing the performance it becomes clear that none of the algorithms performs uniformly better in all conditions. As a result, it is outlined how some of the successful features could be combined to obtain better adaptive MCMC algorithms.
Please contact me via jannik.hartmann "a t" outlook.de if you are interested in the full text.