dsdive.obs.sampleparams_shared: Sample pi and lambda parameters for a single dive stage

In jmhewitt/dsdive: Discrete Space Models for Dive Trajectories of Animals

Description

Sampler uses a Gaussian approximation to the full conditional posterior. Computations are carried out via shared-memory parallelization, and designed for use with the dive-specific covariate model.

Usage

dsdive.obs.sampleparams_shared(
  s0,
  sample.betas,
  theta,
  alpha.priors.list,
  beta.priors.list,
  cl,
  shared.env,
  gapprox = NULL,
  output.gapprox = FALSE,
  rw.sampler = NULL,
  adaptive = FALSE,
  gapprox.approx_ld = FALSE,
  optim.maxit = 1000,
  adaptation.frequency = 10
)

Arguments

s0	the stage for which updated parameters should be sampled
sample.betas	TRUE to sample the directional preference coefficients, or FALSE to sample the speed coefficients.
theta	list containing current values of model parmaeters beta1, beta2, alpha1, alpha2, and alpha3.
alpha.priors.list	List of lists. Each of the sublists specifies the prior distribution for the parameter vectors alpha1, alpha2, and alpha3. See dsdive.gibbs.obs.cov for more details.
beta.priors.list	List of lists. Each of the sublists specifies the prior distribution for the parameter vectors beta1 and beta2. See dsdive.gibbs.obs.cov for more details.
cl	Shared-memory cluster to be used to distribute some computations. The cluster must be fully initialized before running this function. The cluster requires random seeds and Rdsm-initialization.
shared.env	environment containing shared-memory variable pointers. shared.env is expected to be the output from the initialization function gibbs_init_shared.
gapprox	(Optional) gaussapprox object containing the Gaussian approximation used to propose model parameters. If NULL, then a Gaussian approximation will be computed.
output.gapprox	TRUE to return the Gaussian approximation used to propose model parameters.
rw.sampler	If NULL, then a new sampler will be created, otherwise the sampler will be called.
adaptive	TRUE to use adaptive Random walk Metropolis-Hastings proposals instead of Gaussian approximations
gapprox.approx_ld	TRUE to use an approximate likelihood when building Gaussian approximations to full conditional posteriors.
optim.maxit	maximum number of steps to take during numerical optimization to compute Gaussian approximation to full conditional posteriors used to propose model parameters
adaptation.frequency	Random walk proposals will only be updated at intervals of this step count

Examples

## 
## input for gibbs sampler
##

data('dive.sim')
attach(dive.sim)
attach(dive.sim$params)
library(parallel)
library(Rdsm)

cl = makeCluster(2, 'SOCK')
clusterEvalQ(cl, library(dsdive))
clusterEvalQ(cl, library(Rdsm))

mgrinit(cl)

t.stages = sim$times[c(FALSE,diff(sim$stages)==1)]

tstep = diff(sim.obs$times[1:2])

obstx.mat = lapply(1:3, function(s) {
  dsdive.obstx.matrix(depth.bins = depth.bins, beta = beta, 
                      lambda = lambda, s0 = s, tstep = tstep, 
                      include.raw = TRUE, delta = 1e-10)
})

alpha.priors.list = list(
  list(mu = 0, sd = 3),
  list(mu = 0, sd = 3),
  list(mu = 0, sd = 3)
)

beta.priors.list = list(
  list(mu = 0, sd = 3),
  list(mu = 0, sd = 3)
)

T1.prior.params = c(25, .04)
T2.prior.params = c(56, .06)

dsobs.list = list(sim.obs, sim.obs)
t.stages.list = list(t.stages, t.stages)

beta.init = list(
  c(qlogis(.5), 0),
  c(0, 1)
)

alpha.init = c(beta.init, list(c(-1,-1)))

covs = data.frame(x1 = c(.5, 1), x2 = c(0, .3))

pi.formula = ~x1
lambda.formula = ~x1:x2


##
## barebones setup for gibbs sampler internals
##


n = length(dsobs.list)


#
# expand covariate design matrices
#

if(!inherits(pi.formula, 'list')) {
  pi.formula = list(pi.formula, pi.formula)
}

if(!inherits(lambda.formula, 'list')) {
  lambda.formula = list(lambda.formula, lambda.formula, lambda.formula)
}

pi.designs = lapply(pi.formula, function(f) model.matrix(f, covs))
lambda.designs = lapply(lambda.formula, function(f) model.matrix(f, covs))


#
# initialize nodes and shared memory pass-throughs
#

# required variables in environment
beta1.prior = beta.priors.list[[1]]
beta2.prior = beta.priors.list[[2]]
alpha1.prior = alpha.priors.list[[1]]
alpha2.prior = alpha.priors.list[[2]]
alpha3.prior = alpha.priors.list[[3]]
depth.bins = depth.bins
max.width = 100
max.width.offset = 30
t0.prior.params = c(1,1)
tf.prior.params = c(1,1)
delta = 1e-10
offsets = 0
offsets.tf = 0

shared.env = gibbs_init_shared(cl = cl, envir = environment())


#
# initialize sampler output
#

theta = list(beta1 = shared.env$beta1[], beta2 = shared.env$beta2[],
             alpha1 = shared.env$alpha1[], alpha2 = shared.env$alpha2[],
             alpha3 = shared.env$alpha3[])



##
## demo shared parameter sampling
##

# update stage 1 parameters
theta.raw = dsdive.obs.sampleparams_shared(
  s0 = 1, theta = theta, alpha.priors.list = alpha.priors.list,
  beta.priors.list = beta.priors.list, cl = cl, shared.env = shared.env, 
  optim.maxit = 1, sample.betas = TRUE)


##
## demo distributed likelihood evaluation
##

dsdive.obsld_shared(theta = theta, shared.env = shared.env, 
                    cl = cl, s0 = 1, sf = 3)


##
## clean up
##

stopCluster(cl)

detach(dive.sim$params)
detach(dive.sim)

jmhewitt/dsdive documentation built on May 29, 2020, 5:18 p.m.