If you want to perform MCMC (Metropolis-Hastings) to infer discrete values, what are some proposal distributions you can use for this. I can't think of a way to extend the notion of a gaussian proposal to this scenario.
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Discrete samplings problems are, in a sense, much more diverse than their continuous counter-parts. In essence, the problem structure becomes much more important in terms of choosing how you want to move around on the space.
Often used proposals are binary switches ([0,0,1] to [0,1,0] or [0,0,0], for eample), switching neighbours in permutations ([1,2,3,4] -> [1,3,2,4] or [1,2,4,3], for example), uniform choice over the entire space if it is (very) small, etc.
answered Aug 27, 2020 at 10:37
-
$\begingroup$ Very good point, the notion of neighbourhood is much harder to specify in a discrete setting than in a continuous one. $\endgroup$– Xi'anCommented Sep 9, 2020 at 7:50