monty is the
spiritual replacement for mcstate,
which was used for inference of odin.dust models from
2020 to 2024.
Unlike with the new versions of odin (odin2) and dust (dust2), we have
not changed mcstate in place as we no longer felt that the
name was appropriate; mcstate was explicitly about working with state
space models, while monty is about Monte Carlo methods in
general. And Wes has a goat called Monty, so here we are.
Some features have moved out of this package and into
dust2:
The particle filter has moved into dust2, see dust2::dust_filter_create
What was previously a method on the old particle_filter
object has now changed to be a free function:
| Action | mcstate | dust2 |
|---|---|---|
| Allocate | $new() |
dust_filter_create() (1) |
| Run | $run() |
dust_likelihood_run() |
| Stateful run | $run_begin() |
Not supported |
| Get final state | $state() |
dust_likelihood_last_state() |
| Get trajectories | $history() |
dust_likelihood_last_trajectories() |
| Get ODE statistics | $ode_statistics() |
Not yet supported |
| Get intermediate state | $restart_state() |
Not yet supported |
| Get inputs | $inputs |
Not yet supported |
| Change number of threads | set_n_threads() |
Not supported |
dust_unfilter_create for deterministic modelsIn addition, differentiable deterministic models have
dust_likelihood_last_gradient() to get the gradient of the
likelihood at the last point.
In mcstate, we implemented two other sequential Monte Carlo methods: SMC^2 and IF2. Neither of these have been reimplemented yet, but these will appear in dust if and when we do so. Please let us know if you used these!
How these are specified and configured has totally changed, to the point where it is not really meaningful to do a table-based comparison of methods.
A typical use with mcstate looked like this (adapted with simplification from the old SIR models vignette)
filter <- mcstate::particle_filter$new(data = ..., model = ..., compare = ...,
n_particles = ...)
beta <- mcstate::pmcmc_parameter("beta", 0.2, min = 0)
gamma <- mcstate::pmcmc_parameter("gamma", 0.1, min = 0, prior = function(p)
dgamma(p, shape = 1, scale = 0.2, log = TRUE))
proposal_matrix <- diag(0.1, 2)
mcmc_pars <- mcstate::pmcmc_parameters$new(list(beta = beta, gamma = gamma),
proposal_matrix)
control <- mcstate::pmcmc_control(
n_steps,
save_state = TRUE,
save_trajectories = TRUE,
progress = TRUE)
samples <- mcstate::pmcmc(mcmc_pars, filter, control = control)With monty this would look more like:
filter <- dust_filter_create(generator, time_start, data, n_particles)
packer <- monty_packer(c("beta", "gamma"))
model <- dust_likelihood_monty(filter, packer,
save_state = TRUE,
save_trajectories = TRUE)
prior <- monty_dsl({
beta ~ Uniform(0, 100)
gamma ~ Gamma(shape = 1, scale = 0.2)
})
sampler <- monty_sampler_random_walk(diag(0.2, 2))
samples <- monty_sample(model + prior, sampler, n_steps)which contains roughly the same bits of information but in quite a different presentation:
The way that parameters are specified: in
mcstate we had pmcmc_parameter and
pmcmc_parameters, one of which was a function and the other
was a combination that included the sampler’s proposal kernel! This is
all a bit silly. Now: * we use our packer
interface to smooth over the gap between how it is convenient to
represent parameters in your odin model and how you need to represent
them to think about moving around in parameter space * we use the DSL to
specify priors, removing the need to write out distributions, to sample
from these distributions automatically and to prevent forgetting the
log = TRUE on densities.
The pmcmc_control object has gone: this
contained information that really affected the filter (e.g.,
save_state which has moved into the
dust_likelihood_monty function) and the running of the
chain itself.
The control over samplers has moved into a new sampler object: we needed this additional level of control to allow different samplers to be used in different situations (e.g., to allow HMC in the case of a deterministic differentiable density).