1: \begin{abstract}

2: We consider various versions of adaptive Gibbs and

3: Metropolis-within-Gibbs samplers, which update their

4: selection probabilities (and perhaps also their proposal

5: distributions) on the fly during a run, by learning as

6: they go in an attempt to optimise the algorithm.  We

7: present a cautionary example of how even a simple-seeming

8: adaptive Gibbs sampler may fail to converge.  We then

9: present various positive results guaranteeing convergence

10: of adaptive Gibbs samplers under certain conditions.

11: \end{abstract}

12: