Dear R developers,
these days I'm working on some R code for fitting completely generic
Bayesian Hierarchical Models in R, a la OpenBUGS and JAGS.
A key feature of OpenBUGS and JAGS is that they automatically build an
appropriate MCMC sampler from a generic model, specified as a directed
acyclic graph (DAG).
The spirit of my (would-be) implementation is instead more focused on
experimentation and prototyping, i.e. is the user who explicitely
assign samplers for each model variable after specifying the model.
The sampler can be chosed in a set of predefined samplers, as well as
customly specified by the user as an R or C function in a very
flexible way.
Now I have a prototype scheleton implementation (a bounch of R and C
files, together with some base testing scripts) which works at decent
speed (w.r.t. JAGS) on some example models, and I'm writing a
proof-of-concept, reproducible Sweave file about it, to be published
online shortly.
What do you think about it in general?
What do you think about developing an R package of it as a GSoC project?
Best regards,
Antonio, Fabio Di Narzo.