Changelog • mcstate

Particle filters now work with irregular and non-unit spaced time series data

Continuous time (ODE) models can now use workers for running chains in parallel with pmcmc

Basic inference now working with continuous time (ODE) models, via particle_filter and pmcmc

Support for adaptive proposals for deterministic models

Allow running a particle filter with multiple parameter sets and a single data set.
The nested field on the particle filter class has been split into two logical fields: has_multiple_parameters and has_multiple_data

Deprecated ‘discrete’ argument to parameters in favour of ‘integer’ - affects if2_parameter, pmcmc_parameter, pmcmc_varied_parameter, smc2_parameter

Compiled compare functions now supported in more places - particle_deterministic and multistage models (#177)

Overhaul mcstate::pmcmc_chains_* to always use a file for communication, making them easier to understand and more robust (#179)
New functions mcstate::pmcmc_chains_cleanup for removing files created by the above, and mcstate::pmcmc_chains_collect for automating collecting samples
New, simpler, approach to pmcmc parallelisation which shares as much code with the above.

Allow filtering of the pmcmc chains during running (dropping burnin and filtering) to reduce memory usage when collectin large trajectories
pmcmc no longer retains the initial parameter values

New argument to mcstate::particle_filter and mcstate::particle_deterministic, constant_log_likelihood which can be used to compute the probabilities of non-time series data (#185)

Rework the “nested” support; this now returns output in a different dimension order. Primarily this is an internal refactoring.
Allow use of multistage parameters with deterministic models, and with nested parameters.
Transform functions for multistage parameters now take info and not model as an argument, more in keeping with other functions.

Allow multistage parameters to work with the “deterministic” particle
New mcstate::particle_deterministic_state object for advanced use of the deterministic particle
Deterministic particle loses the run_many method

Multistage particle filters now cope with running data covering a subset of their stages
Drop support for chnging initial step via particle filter initial function for deterministic and nested filters

New helper function mcstate::particle_filter_initial for creating particle filter initial state functions from restart data.
Drop support for changing initial step via the particle filter initial function

Multi-stage particle filter implemented, allowing arbitrary changes to model structure during a particle filter run (#159)

Allow saving restart from the deterministic filter (#153)

Reduced overhead in parallel pmcmc with workers, and faster/less memory-hungry chain combination (#142)

Allow the particle filter to terminate early if we would not be interested in the result. This is useful for mcstate::pmcmc which can use it to stop calculating a likelood that would be rejected. Primarily useful when running with relatively low numbers of particles and a high variance in the estimator (#138)

Add support for running in “deterministic” mode with recent dust (#139)

Add an iterated filtering method via mcstate::if2 (#123)

New functions pmcmc_chains_prepare and pmcmc_chains_run which can be used to manually schedule chains over different computing resourcess (#129)

When rerun_every is specified, a new control parameter rerun_control can be used to make this stochastic rerun

The particle filter can now run entirely in compiled code if supported by the model. This may give a small performance gain, particularly on very simple models, or of the model has an expensive compare function (#118)

Add nested_step_ratio parameter to pmcmc_control for controlling the ratio of fixed:varied steps for nested pMCMC

New array helper mcstate::array_flatten for unshaping an array

Remove deprecated arguments to pmcmc (these were deprecated in 0.3.0) (#114)

Bugfix in index for nested particle filters.

Extend support of pmcmc to pmcmc_parameters_nested objects.

Added particle_filter_state_nested and extended particle_filter to handle pmcmc_parameters_nested objects.

Basic SMC^2 implementation (smc2()) as an alternative to pmcmc. This is very embryonic and the interface will change over future versions to support things like restarting and saving trajectories (#13)

Bug fixes in $proposal method of pmcmc_parameters_nested for discrete and bounded parameters.

Added helper methods mcstate::array_bind, mcstate::array_reshape and mcstate::array_drop to simplify some common array operations (#106)

Added pmcmc_varied_parameter for parameters that can vary between different populations.
Added pmcmc_parameters_nested to hold parameters that vary between populations (pmcmc_varied_parameter) and parameters that are the same (fixed) between populations (pmcmc_parameter).

Fix performance regression added in 0.4.3

Support for incrementally running a particle filter (up to some point in the time series) and forking these partial runs; see the $begin_run method on the particle filter (#78)

Fix typo in sir_models.Rmd

New $fix() method on pmcmc_parameters objects for fixing the value for a subset of parameters before running with pmcmc (#98)

Compare functions no longer use (or accept) the prev_state argument and now use just the current model state. This requires that models compute things like “daily incidence” within model code but simplifies use with irregular time series (#94)

Support for “compiled compare functions”, introduced in dust 0.6.1 (#92)

The particle filter can now return the entire model state at points during the run, with argument save_restart to $run() and method $restart_state() (#86)
The pmcmc can returned sample restart state using the save_restart argument to mcstate::pmcmc_control which can be used to restart the pMCMC part way along the time series (see vignette("restart"))

pmcmc is now controllable via a new mcstate::pmcmc_control object
pmcmc can run chains in parallel using callr, by specifying n_workers = n for n greater than 1.

pmcmc adds new rerun_every argument to rerun the particle filter unconditionally.