Dual-averaging for other MCMC algorithms



This not a Stan question, but as have many others, I’m looking to take advantage of the remarkable concentration of smart people on this forum. Looking at you, @Bob_Carpenter and @betanalpha .

Would it (in principle) pay off to implement primal-dual averaging as the adaptation scheme of other MCMC packages/algorithms? Is there any reason why one cannot use primal dual-averaging to set the scaling \sigma in, say, RW Metropolis-Hastings? Has anyone seen anything in that direction?



I don’t know your exact problem that you want to solve, but you could start by having a look at “A tutorial on adaptive MCMC” by Andrieu and Thoms. Paper Link


Thanks, I’ve read that one. It seems to me that primal-dual averaging does attain the regularity conditions for diminishing adaptation discussed by Andrieu and Thoms, no? My question is more in the sense of whether it’d pay to implement something like primal-dual averaging in, say, any of the adaptive MCMC packages in R.