configuring sampler for a subset of elements of an array

[Torbjørn Ergon]

Hi all,

I'm working a a spatial model. The default samplers are: 

posterior_predictive sampler (2301) - z[] (2288 elements) - x[] (13 multivariate elements) binary sampler (9882) - z.cam[] (440 elements) - z[] (9442 elements) RW_block sampler (85) - x[] (85 multivariate elements) RW sampler (77)

...

I get very bad mixing and want to adjust 'adaptFactorExponent' for the RW_block sampler. x[] is an array of the spatial random effects for 138 sites, 14 occasions, and 7 species. Because some site-occasion combinations are not sampled, some elements of x[] have a posterior_predictive sampler. It will be extremely complicated to avoid this since we are not allowed to use vector indices to vectors and matrices (for dmnorm()). The following fails,

conf$removeSamplers('x[1:138, 1:14, 1:7]')
conf$addSampler('x[1:138, 1:14, 1:7]', 'RW_block', control = list(adaptFactorExponent = 0.25))

because that removes the  posterior_predictive samplers.

Is there an easy way to access only the nodes in x[] that have the 'RW_block' sampler?

Cheers,

Torbjørn

[Daniel Turek]

When you first call configureMCMC, try passing the control list argument as:

conf <- configureMCMC(model, ..., control = list(adaptFactorExponent = 0.25))

I believe that will pass the control argument "adaptFactorExponent = 0.25" to all samplers which are assigned, so that should catch all the RW_block samplers which are assigned by default (without affecting the predictive samplers).

Cheers,

Daniel

--
You received this message because you are subscribed to the Google Groups "nimble-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nimble-users...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/nimble-users/1f8d1a2f-5ae4-4a7f-bbd8-1d204ffc4e42n%40googlegroups.com.

[Torbjørn Ergon]

Thanks Daniel! That worked!

Still poor mixing though. I.e., low autocorrelation within chains and each chain seems stationary, but they converge to different distributions. Not sure what that indicates?

Torbjørn

[Chris Paciorek]

You might try the HMC sampler or the "Barker" sampler.

To view this discussion visit https://groups.google.com/d/msgid/nimble-users/a3800473-9b0f-47e4-b87b-386c076cca8an%40googlegroups.com.