NEWS.md
roxygen2 issue #1491
Fixed issue #21: Restricting search scope in MCMC temp environment data.
Removed annoying warning when using convergence checker.
The function fun passed to MCMC is now called two times less. It shouldn’t significantly affect any previous results.
convergence_gelman now stores the Gelman and Rubin’s statistics in the correct order, i.e., the most recent at the end of the array in convergence_data_get("val").
Users can now pass seed to MCMC. If is.null(seed) != TRUE, then seed is passed to set.seed().
The function convergence_auto() now behaves as expected. Before, it was not checking convergence.
The set of functions last_* and LAST_MCMC will be deprecated in favor of get_* and MCMC_OUTPUT.
The new function get_logpost() returns the computed values of the objective function from the last MCMC run.
The new function get_draws() returns the MCMC draws from the kernel’s proposal function (proposed states).
The new function set_userdata(...) allows storing information into a data.frame as the MCMC process runs. Users can retrieve the data with the function get_userdata().
The new function ith_step() provides access to objects within the MCMC loop during the run. The new function comes with a vignette that illustrates its usage.
The function append_chains() was randomly dropping one sample of the final set.
A new artificial dataset lifeexpect is shipped with the package. This simulates 1,000 observations of age at death using US’s statistics.
kernel_am and kernel_ram no longer fail when at least one parameter is an offset (fixed = TRUE for some parameter).
Now kernel_ram tries first to find the cholesky decomp. If it fails, then it uses Matrix::nearPD and re-tries. This is following what is done in the adaptMCMC package.
Workflow for running MCMC with conv_checker re-designed (less error prone).
Environments LAST_RUN and LAST_CONV_CHECK provide information about the last call to MCMC and the corresponding convergence checker. Users can access these environments via getter and setter functions.
MCMC with convergence checker now reports the status of the convergence statistic using the LAST_CONV_CHECK environment and corresponding functions.
The functions to compute mean and variance recursively now allow us to do so using windows.
Adding Vihola (2012)’s Robust Adaptive Metropolis, Haario et al. (2001)’s Adaptive Metropolis, and Thawornwattana et al. (2018)’s mirror kernel algorithms.
The argument progress is no longer ignored. When set to TRUE, the function will print the progress of the MCMC algorithm.
Improved coverage and fixed minor bugs.
When running with convergence check, fixed parameters (offset), as tagged in the fmcmc_kernel object will be excluded from the call to conv_checker.