• 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
  • 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)
  • 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
  • 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 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.