rng_seed appears to no effect in map2stan()

[rmcelreath]

rng_seed appears to no effect in map2stan()

[rmcelreath]

Yeah, this is a mysterious issue. The map2stan argument is rng_seed. It's passed to stan as seed but seems not to have an effect. No idea why, since passing a value in a raw stan call works. But I can't yet figure out a way that my code is the issue. You can even check the internal stanfit object and see that stan is saving the passed seed. It just doesn't appear to use it?

Maybe something will occur to me. In the meantime, just use set.seed before the map2stan call.

[kmiddleton]

It might be the init parameter in lines 1356 and 1358 of map2stan.R. The comment in this SO question suggests that seed only covers the RNG from the Boost library, and that init draws separately.

set.seed() coupled with set_rng does appear to work:

library(rethinking)
data("Howell1")

d <- Howell1
d2 <- d[ d$age >= 18 , ]

set.seed(5)

m1 <- map2stan(
  alist(
    height ~ dnorm(mu, sigma),
    mu ~ dnorm(178, 20),
    sigma ~ dunif(0, 50)),
  data = d2,
  WAIC = FALSE,
  rng_seed = 8
)

set.seed(5)

m2 <- map2stan(
  alist(
    height ~ dnorm(mu, sigma),
    mu ~ dnorm(178, 20),
    sigma ~ dunif(0, 50)),
  data = d2,
  WAIC = FALSE,
  rng_seed = 8
)

m1
m2

all.equal(m1, m2)

The only differences are in the date, time, dso, and model_cpp.

[jgabry]

On Tuesday, August 30, 2016, Kevin Middleton [email protected]
wrote:

The comment in this SO question
http://stackoverflow.com/questions/27204096/supplying-seed-to-stan-doesnt-guarantee-the-same-chains
suggests that seed only covers the RNG from the Boost library, and that
init draws separately.

Yes, that's correct. If any initial value generation is being done in R
(whether by you or by passing a function to init) then you'll have to set
both the R seed and the seed for the Boost library.

[rmcelreath]

Thanks. Sounds like an easy fix, with this information. Will try to roll it into the next push.

[rmcelreath]

This is fixed in experimental branch.