initializing parallel chains in rjags - r

Question

I'm doing some ghetto parallelization in jags through rjags.
I've been using the function parallel.seeds to obtain RNG states to intialize the RNG's (example below). However, I don't understand why multiple integers are returned for each RNG. In the documentation it says that when you intialize .RNG.state is supposed to be a numeric vector with length one.
Furthermore, sometimes when I try to do this R crashes with no error generated. When I give up and just let it generate the seed for the chain on it's own, the model runs fine. Does this mean I am using the wrong .RNG.state? Any insight would be appreciated, as I am planning to scale up this model in the future.
> parallel.seeds("base::BaseRNG", 3)
[[1]]
[[1]]$.RNG.name
[1] "base::Wichmann-Hill"
[[1]]$.RNG.state
[1] 3891 16261 19841
[[2]]
[[2]]$.RNG.name
[1] "base::Marsaglia-Multicarry"
[[2]]$.RNG.state
[1] 408065014 1176110892
[[3]]
[[3]]$.RNG.name
[1] "base::Super-Duper"
[[3]]$.RNG.state
[1] -848274653 175424331

Answer 1

There is a difference between .RNG.seed (which is a vector of length one, and the thing you can specify to jags.model to e.g. ensure MCMC samples are repeatable) and .RNG.state (which is a vector of length depending on the pRNG algorithm). It is possible that these got mixed up in the docs somewhere - can you tell me where you read this so I can make sure it is fixed for JAGS/rjags 4?
Regarding the crashing - some more details would be needed to help you with that I'm afraid. I assume that it is the JAGS model that crashes, and not your R session that terminates, and after the model has been running for a while? A reproducible example would help a lot.
By the way - when you say 'scale up' - if you are planning to make use of > 4 chains I would strongly recommend you load the lecuyer module (see ?parallel.seeds examples at the bottom).
Matt

Answer 2

The documentation is a bit confusing; under ?jags.model we see that .RNG.seed should be a vector of length 1, but parallel.seeds() returns .RNG.state which is usually > 1. The state space for the Mersenne Twister algorithm has 624 integers, and that is the length of the vector when you do
parallel.seeds("base::BaseRNG",4)
to make sure you see all 4 types of RNG. Similarly the state space of the Wichmann-Hill generator has 3 integers, and I'm sure similar research would reveal the state spaces for the other two are longer than 1.
For my own edification I mocked up an example using the LINE data in rjags:
data(LINE)
LINE$model() ## edit and save to line.r
data = LINE$data()
line = jags.model("line.r",data=data)
line.samples <- jags.samples(LINE, c("alpha","beta","sigma"),n.iter=1000)
line.samples
inits = parallel.seeds("base::BaseRNG", 3) # a list of lists
inits[[1]]$tau = 1
inits[[1]]$alpha = 3
inits[[1]]$beta = 1
inits[[2]]$tau = .1
inits[[2]]$alpha = .3
inits[[2]]$beta = .1
inits[[3]]$tau = 10
inits[[3]]$alpha = 10
inits[[3]]$beta = 5
line = jags.model("line.r",data=data,inits=inits,n.chains=3)
line.samples <- jags.samples(line, c("alpha","beta","sigma"),n.iter=1000)
line2 = jags.model("line.r",data=data,inits=inits,n.chains=3)
line.samples2 <- jags.samples(line2, c("alpha","beta","sigma"),n.iter=1000)
all(line.samples$alpha-line.samples2$alpha < 0.00000001) ## TRUE
So the results are entirely repeatable, which is cool.
To understand the conditions under which R is crashing, I'd need to know the results of sessionInfo() on your computer, plus more details of the circumstances (e.g. what JAGS model are you running?). I just did:
for (i in 1:100){parallel.seeds("base::BaseRNG",4)}
and my computer didn't crash. For reference:
sessionInfo()
# R version 3.1.3 (2015-03-09)
# Platform: x86_64-w64-mingw32/x64 (64-bit)
# Running under: Windows 7 x64 (build 7601) Service Pack 1
#
# locale:
# [1] LC_COLLATE=English_United States.1252
# [2] LC_CTYPE=English_United States.1252
# [3] LC_MONETARY=English_United States.1252
# [4] LC_NUMERIC=C
# [5] LC_TIME=English_United States.1252
#
# attached base packages:
# [1] stats graphics grDevices utils datasets
# [6] methods base
#
# other attached packages:
# [1] rjags_3-14 coda_0.17-1 mlogit_0.2-4
# [4] maxLik_1.2-4 miscTools_0.6-16 Formula_1.2-1
#
# loaded via a namespace (and not attached):
# [1] grid_3.1.3 lattice_0.20-30 lmtest_0.9-33
# [4] MASS_7.3-39 sandwich_2.3-3 statmod_1.4.21
# [7] tools_3.1.3 zoo_1.7-12
That shows the version of R and rjags that I'm using.