dHMM: Hidden Markov Model distribution for use in 'nimble' models

In nimbleEcology: Distributions for Ecological Models in 'nimble'

Hidden Markov Model distribution for use in nimble models

Description

dHMM and dHMMo provide hidden Markov model distributions that can be used directly from R or in nimble models.

Usage

dHMM(x, init, probObs, probTrans, len = 0, checkRowSums = 1, log = 0)

dHMMo(x, init, probObs, probTrans, len = 0, checkRowSums = 1, log = 0)

rHMM(n, init, probObs, probTrans, len = 0, checkRowSums = 1)

rHMMo(n, init, probObs, probTrans, len = 0, checkRowSums = 1)

Arguments

x	vector of observations, each one a positive integer corresponding to an observation state (one value of which could can correspond to "not observed", and another value of which can correspond to "dead" or "removed from system").
init	vector of initial state probabilities. Must sum to 1
probObs	time-independent matrix (dHMM and rHMM) or time-dependent array (dHMMo and rHMMo) of observation probabilities. First two dimensions of probObs are of size x (number of possible system states) x (number of possible observation classes). dDHMMo and rDHMMo expects an additional third dimension of size (number of observation times). probObs[i, j (,t)] is the probability that an individual in the ith latent state is recorded as being in the jth detection state (at time t). See Details for more information.
probTrans	time-independent matrix of state transition probabilities. probTrans[i,j] is the probability that an individual in latent state i transitions to latent state j at the next timestep. See Details for more information.
len	length of x (see below).
checkRowSums	should validity of probObs and probTrans be checked? Both of these are required to have each set of probabilities sum to 1 (over each row, or second dimension). If checkRowSums is non-zero (or TRUE), these conditions will be checked within a tolerance of 1e-6. If it is 0 (or FALSE), they will not be checked. Not checking should result in faster execution, but whether that is appreciable will be case-specific.
log	TRUE or 1 to return log probability. FALSE or 0 to return probability.
n	number of random draws, each returning a vector of length len. Currently only n = 1 is supported, but the argument exists for standardization of "r" functions.

Details

These nimbleFunctions provide distributions that can be used directly in R or in nimble hierarchical models (via nimbleCode and nimbleModel).

The distribution has two forms, dHMM and dHMMo. Define S as the number of latent state categories (maximum possible value for elements of x), O as the number of possible observation state categories, and T as the number of observation times (length of x). In dHMM, probObs is a time-independent observation probability matrix with dimension S x O. In dHMMo, probObs is a three-dimensional array of time-dependent observation probabilities with dimension S x O x T. The first index of probObs indexes the true latent state. The second index of probObs indexes the observed state. For example, in the time-dependent case, probObs[i, j, t] is the probability at time t that an individual in state i is observed in state j.

probTrans has dimension S x S. probTrans[i, j] is the time-independent probability that an individual in state i at time t transitions to state j time t+1.

init has length S. init[i] is the probability of being in state i at the first observation time. That means that the first observations arise from the initial state probabilities.

For more explanation, see package vignette (vignette("Introduction_to_nimbleEcology")).

Compared to writing nimble models with a discrete latent state and a separate scalar datum for each observation time, use of these distributions allows one to directly sum (marginalize) over the discrete latent state and calculate the probability of all observations for one individual (or other HMM unit) jointly.

These are nimbleFunctions written in the format of user-defined distributions for NIMBLE's extension of the BUGS model language. More information can be found in the NIMBLE User Manual at https://r-nimble.org.

When using these distributions in a nimble model, the left-hand side will be used as x, and the user should not provide the log argument.

For example, in nimble model code,

observedStates[i, 1:T] ~ dHMM(initStates[1:S], observationProbs[1:S, 1:O], transitionProbs[1:S, 1:S], 1, T)

declares that the observedStates[i, 1:T] (observation history for individual i, for example) vector follows a hidden Markov model distribution with parameters as indicated, assuming all the parameters have been declared elsewhere in the model. As above, S is the number of system state categories, O is the number of observation state categories, and T is the number of observation occasions. This will invoke (something like) the following call to dHMM when nimble uses the model such as for MCMC:

dHMM(observedStates[1:T], initStates[1:S], observationProbs[1:S, 1:O], transitionProbs[1:S, 1:S], 1, T, log = TRUE)

If an algorithm using a nimble model with this declaration needs to generate a random draw for observedStates[1:T], it will make a similar invocation of rHMM, with n = 1.

If the observation probabilities are time-dependent, one would use:

observedStates[1:T] ~ dHMMo(initStates[1:S], observationProbs[1:S, 1:O, 1:T], transitionProbs[1:S, 1:S], 1, T)

Value

For dHMM and dHMMo: the probability (or likelihood) or log probability of observation vector x.

For rHMM and rHMMo: a simulated detection history, x.

Notes for use with automatic differentiation

The dHMM[o] distributions should work for models and algorithms that use nimble's automatic differentiation (AD) system. In that system, some kinds of values are "baked in" (cannot be changed) to the AD calculations from the first call, unless and until the AD calculations are reset. For the dHMM[o] distributions, the sizes of the inputs and the data (x) values themselves are baked in. These can be different for different iterations through a for loop (or nimble model declarations with different indices, for example), but the sizes and data values for each specific iteration will be "baked in" after the first call. In other words, it is assumed that x are data and are not going to change.

Author(s)

Ben Goldstein, Perry de Valpine, and Daniel Turek

References

D. Turek, P. de Valpine and C. J. Paciorek. 2016. Efficient Markov chain Monte Carlo sampling for hierarchical hidden Markov models. Environmental and Ecological Statistics 23:549–564. DOI 10.1007/s10651-016-0353-z

See Also

For dynamic hidden Markov models with time-dependent transitions, see dDHMM and dDHMMo. For simple capture-recapture, see dCJS.

Examples

# Set up constants and initial values for defining the model
len <- 5 # length of dataset
dat <- c(1,2,1,1,2) # A vector of observations
init <- c(0.4, 0.2, 0.4) # A vector of initial state probabilities
probObs <- t(array( # A matrix of observation probabilities
       c(1, 0,
         0, 1,
         0.2, 0.8), c(2, 3)))
probTrans <- t(array( # A matrix of transition probabilities
        c(0.6, 0.3, 0.1,
          0, 0.7, 0.3,
          0, 0, 1), c(3,3)))

# Define code for a nimbleModel
 nc <- nimbleCode({
   x[1:5] ~ dHMM(init[1:3], probObs = probObs[1:3,1:2],
                 probTrans = probTrans[1:3, 1:3], len = 5, checkRowSums = 1)

   for (i in 1:3) {
     for (j in 1:3) {
       probTrans[i,j] ~ dunif(0,1)
     }

     probObs[i, 1] ~ dunif(0,1)
     probObs[i, 2] <- 1 - probObs[i, 1]
   }
 })

# Build the model
HMM_model <- nimbleModel(nc,
                         data = list(x = dat),
                         inits = list(init = init,
                                      probObs = probObs,
                                      probTrans = probTrans))
# Calculate log probability of data from the model
HMM_model$calculate()
# Use the model for a variety of other purposes...

nimbleEcology documentation built on June 27, 2024, 5:09 p.m.