mcstate_model()
Create basic model
mcstate_sample()
Sample from a model
mcstate_sampler_random_walk()
Random Walk Sampler
R/model.R
mcstate_model.Rd
Create a basic mcstate model. Currently nothing here is +validated, and it's likely that users will never actually use this +directly. Contains data and methods that define a basic model +object, so that we can implement samplers against. Not all models +will support everything here, and we'll add additional +fields/traits over time to advertise what a model can do. For +example, models will need to advertise that they are capable of +being differentiated, or that they are stochastic in order to be +used with different methods.
mcstate
mcstate_model(parameters, direct_sample, density, gradient, domain)
Names of the parameters. Every parameter is +named, and for now every parameter is a scalar. We might relax +this later to support an odin-style structured parameter list, +but that might just generate a suitable vector of parameter +names perhaps? In any case, once we start doing inference it's +naturally in the R^n, and here n is defined as the length of +this vector of names.
odin
A function to sample directly from the +parameter space. In the case where a model returns a posterior +(e.g., in Bayesian inference), this is assumed to be sampling +from the prior. We'll use this for generating initial +conditions for MCMC where those are not given, and possibly +other uses.
Compute the model density for a vector of parameter +values; this is the posterior probability in the case of +Bayesian inference, but it could be anything really. Models can +return -Inf if things are impossible, and we'll try and cope +gracefully with that wherever possible.
-Inf
Compute the gradient of density with respect to +the parameter vector; takes a parameter vector and returns a +vector the same length. For efficiency, the model may want to +be stateful so that gradients can be efficiently calculated +after a density calculation, or density after gradient, where +these are called with the same parameters.
density
Information on the parameter domain. This is a two +column matrix with length(parameters) rows representing each +parameter. The parameter minimum and maximum bounds are given +as the first and second column. Infinite values (-Inf or +Inf) should be used where the parameter has infinite domain up +or down. Currently used to translate from a bounded to +unbounded space for HMC, but we might also use this for +reflecting proposals in MCMC too.
length(parameters)
Inf
An object of class mcstate_model, which can be used with +a sampler.
mcstate_model
R/sample.R
mcstate_sample.Rd
Sample from a model. Uses a Monte Carlo method (or possibly +something else in future) to generate samples from your +distribution. This is going to change a lot in future, as we add +support for distributing over workers, and for things like +parallel reproducible streams of random numbers. For now it just +runs a single chain as a proof of concept.
mcstate_sample(model, sampler, n_steps, initial = NULL)
The model to sample from; this should be a +mcstate_model for now, but we might change this in future to +test to see if things match an interface rather than a +particular class attribute.
A sampler to use. These will be described later, +but we hope to make these reasonably easy to implement so that +we can try out different sampling ideas. For now, the only +sampler implemented is mcstate_sampler_random_walk().
The number of steps to run the sampler for.
Optionally, initial parameter values for the +sampling. If not given, we sample from the model (or its prior).
A list of parameters and densities.
R/sampler-random-walk.R
mcstate_sampler_random_walk.Rd
Create a simple random walk sampler, which uses a symmetric +proposal to move around parameter space.
mcstate_sampler_random_walk(proposal = NULL, vcv = NULL)
A proposal function; must take a vector of +parameters and produce a new vector of proposed parameters.
A variance covariance matrix to generate a proposal +function from. If you want multivariate Gaussian proposal, this +is likely simpler than supplying your own proposal, and +generally more efficient too.
proposal
A mcstate_sampler object, which can be used with +mcstate_sample
mcstate_sampler