Skip to content

NEWS.md

Latest commit

 

History

History
263 lines (192 loc) · 12.1 KB

NEWS.md

File metadata and controls

263 lines (192 loc) · 12.1 KB
Raw

hBayesDM 2.0.0 (in development)

Major refactor; breaking changes.

Backend swap: rstan → cmdstanr

  • The Stan backend is now CmdStan via cmdstanr, replacing rstan. CmdStan is a system dependency, not a CRAN package — install it once with cmdstanr::install_cmdstan() after installing cmdstanr from the Stan r-universe (see README).
  • The fit slot on the result object is now a CmdStanMCMC object (or CmdStanVB when vb = TRUE), no longer a rstan::stanfit. Methods that used to apply directly (e.g. rstan::extract(fit)) need to be replaced with the cmdstanr / posterior equivalents — fit$draws(), fit$summary(), posterior::as_draws_*(), etc.
  • Stan files have been canonicalized to the modern syntax (array[N, T] real x instead of real x[N, T]; abs replacing fabs).

Tooling and prerequisites

  • R ≥ 4.4 is now required.
  • hBayesDM no longer compiles Stan models at install time. Each model compiles on first use (~30 s) and cmdstanr caches the binary for subsequent fits. The BUILD_ALL install-time flag is gone.
  • LinkingTo: rstan, StanHeaders, ... and the C++ machinery they entailed have been removed from the package.

API changes

  • rhat() — internally uses posterior::rhat via fit$summary() (a workaround for a name-collision bug between hBayesDM::rhat and the string "rhat" looked up by cmdstanr::CmdStanFit$summary()).
  • extract_ic() — now extracts log_lik from fit$draws() via posterior::as_draws_array(). The ic = "looic" | "waic" | "both" API is unchanged.
  • plot.hBayesDM() and plot_ind() — now plot via bayesplot (mcmc_trace, mcmc_intervals, mcmc_areas) instead of rstan::stan_plot.
  • Breaking — snake_case sweep. The R public API is now fully snake_case, matching modern R style (tidyverse / Google guides) and the Python package character-for-character. User-visible renames:
    • Functions: printFit → print_fit, plotHDI → plot_hdi, plotInd → plot_ind, plotDist → plot_dist, HDIofMCMC → hdi.
    • Arguments: modelRegressor → model_regressor, indPars → ind_pars, credMass → ci_prob (on hdi() and plot_hdi()), fontSize → font_size, xLab → x_lab, yLab → y_lab, xLim → x_lim, binSize → bin_size, sampleVec → sample_vec, roundTo → round_to, Title → title.
    • Result-object slots: $allIndPars → $all_ind_pars, $parVals → $par_vals, $modelRegressor → $model_regressor, $rawdata → $raw_data. ($model and $fit unchanged.)
  • HDI default credible mass. hdi() and plot_hdi() default ci_prob = 0.95 (unchanged on the R side). The Python package now also defaults to 0.95 (previously 0.94, inherited from arviz) so the two languages produce matching HDI bands out of the box.
  • additional_args plumbing fixed: model wrappers that declared a NULL default (e.g. banditNarm_2par_lapse Narm, pstRT_ddm initQ) were silently dropping it because args[[nm]] <- NULL removes a list element in R; now uses args[nm] <- list(NULL) to preserve the entry.

Migration notes

If you have downstream code that calls rstan::extract(output$fit), replace it with one of:

posterior::as_draws_df(output$fit$draws())   # tidy data.frame
posterior::as_draws_rvars(output$fit$draws())
output$par_vals[["mu_k"]]                     # already-extracted samples

hBayesDM 1.3.1

  • Add plot functions for Hierarchical Gaussian Filter models: plot_hgf_ibrb, plot_hgf_ibrb_single.

hBayesDM 1.3.0

hBayesDM 1.2.1

  • Fixed a pkgdown error.

hBayesDM 1.2.0

  • Added a drift diffusion model and two reinforcement learning-drift diffision models for the probabilistic selection task: pstRT_ddm, pstRT_rlddm1, and pstRT_rlddm6.
  • Added multiple models for the banditNarm task: banditNarm_2par_lapse, banditNarm_4par, banditNarm_delta, banditNarm_kalman_filter, banditNarm_lapse, banditNarm_lapse_decay, and banditNarm_singleA_lapse.
  • Fixed bart_ewmv to avoid dividing by zero.

hBayesDM 1.1.1

  • Fix symbolic link errors for stan files and example data.

hBayesDM 1.1.0

  • Added the cumulative model for the Cambridge gambling task: cgt_cm.
  • Added two new models for aversive learning tasks: alt_delta and alt_gamma.
  • Added exponential-weight mean-variance model for BART task: bart_ewmv.
  • Added simple Q learning model for the probabilistic selection task: prl_Q.
  • Added signal detection theory model for 2-alternative forced choice task: task2AFC_sdt.

hBayesDM 1.0.2

  • Fixed an error on using data.frame objects as data (#112).

hBayesDM 1.0.1

  • Minor fix on the plotting function.

hBayesDM 1.0.0

Major changes

  • Now, hBayesDM has both R and Python version, with same models included! You can run hBayesDM with a language you prefer!
  • Models in hBayesDM are now specified as YAML files. Using the YAML files, R and Python codes are generated automatically. If you want to contribute hBayesDM by adding a model, what you have to do is just to write a Stan file and to specify its information! You can find how to do in the hBayesDM wiki (https://github.com/CCS-Lab/hBayesDM/wiki).
  • Model functions try to use parameter estimates using variational Bayesian methods as its initial values for MCMC sampling by default (#96). If VB estimation fails, then it uses random values instead.
  • The data argument for model functions can handle a data.frame object (#2, #98).
  • choiceRT_lba and choiceRT_lba_single are temporarily removed since their codes are not suitable to the new package structure. We plan to re-add the models in future versions.
  • The Cumulative Model for Cambridge Gambling Task is added (cgt_cm; #108).

Minor changes

  • The tau parameter in all models for the risk aversion task is modified to be bounded to [0, 30] (#77, #78).
  • bart_4par is fixed to compute subject-wise log-likelihood (#82).
  • extract_ic is fixed for its wrong rep function usage (#94, #100).
  • The drift rate (delta parameter) in choiceRT_ddm and choiceRT_ddm_single is unbounded and now it is estimated between [-Inf, Inf] (#95, #107).
  • Fix a preprocessing error in choiceRT_ddm and choiceRT_ddm_single (#95, #109).
  • Fix igt_orl for a wrong Matt trick operation (#110).

hBayesDM 0.7.2

  • Add three new models for the bandit4arm task: bandit4arm_2par_lapse, bandit4arm_lapse_decay and bandit4arm_singleA_lapse.
  • Fix various (minor) errors.

hBayesDM 0.7.1

  • Make it usable without manually loading rstan.
  • Remove an annoying warning about using ..insensitive_data_columns.

hBayesDM 0.7.0

  • Now, in default, you should build a Stan file into a binary for the first time to use it. To build all the models on installation, you should set an environmental variable BUILD_ALL to true before installation.
  • Now all the implemented models are refactored using hBayesDM_model function. You don't have to change anything to use them, but developers can easily implement new models now!
  • We added a Kalman filter model for 4-armed bandit task (bandit4arm2_kalman_filter; Daw et al., 2006) and a probability weighting function for general description-based tasks (dbdm_prob_weight; Erev et al., 2010; Hertwig et al., 2004; Jessup et al., 2008).
  • Initial values of parameter estimation for some models are updated as plausible values, and the parameter boundaries of several models are fixed (see more on issue #63 and #64 in Github).
  • Exponential and linear models for choice under risk and ambiguity task now have four model regressors: sv, sv_fix, sv_var, and p_var.
  • Fix the Travix CI settings and related codes to be properly passed.

hBayesDM 0.6.3

  • Update the dependencies on rstan (>= 2.18.1)
  • No changes on model files, as same as the version 0.6.2

hBayesDM 0.6.2

  • Fix an error on choiceRT_ddm (#44)

hBayesDM 0.6.1

  • Solve an issue with built binary files.
  • Fix an error on peer_ocu with misplaced parentheses.

hBayesDM 0.6.0

  • Add new tasks (Balloon Analogue Risk Task, Choice under Risk and Ambiguity Task, Probabilistic Selection Task, Risky Decision Task (a.k.a. Happiness task), Wisconsin Card Sorting Task)
  • Add a new model for the Iowa Gambling Task (igt_orl)
  • Change priors (Half-Cauchy(0, 5) --> Half-Cauchy(0, 1) or Half-Normal(0, 0.2)
  • printFit function now provides LOOIC weights and/or WAIC weights

hBayesDM 0.5.1

  • Add models for the Two Step task
  • Add models without indecision point parameter (alpha) for the PRL task (prl_*_woa.stan)
  • Model-based regressors for the PRL task are now available
  • For the PRL task & prl_fictitious.stan & prl_fictitious_rp.stan --> change the range of alpha (indecision point) from [0, 1] to [-Inf, Inf]

hBayesDM 0.5.0

  • Support variational Bayesian methods (vb=TRUE)
  • Allow posterior predictive checks, except for drift-diffusion models (inc_postpred=TRUE)
  • Add the peer influence task (Chung et al., 2015, USE WITH CAUTION for now and PLEASE GIVE US FEEDBACK!)
  • Add 'prl_fictitious_rp' model
  • Made changes to be compatible with the newest Stan version (e.g., // instead of # for commenting).
  • In 'prl_*' models, 'rewlos' is replaced by 'outcome' so that column names and labels would be consistent across tasks as much as possible.
  • Email feature is disabled as R mail package does not allow users to send anonymous emails anymore.
  • When outputs are saved as a file (*.RData), the file name now contains the name of the data file.

hBayesDM 0.4.0

  • Add a choice reaction time task and evidence accumulation models
    • Drift diffusion model (both hierarchical and single-subject)
    • Linear Ballistic Accumulator (LBA) model (both hierarchical and single-subject)
  • Add PRL models that can fit multiple blocks
  • Add single-subject versions for the delay discounting task (dd_hyperbolic_single and dd_cs_single).
  • Standardize variable names across all models (e.g., rewlos --> outcome for all models)
  • Separate versions for CRAN and GitHub. All models/features are identical but the GitHub version contains precompilled models.

hBayesDM 0.3.1

  • Remove dependence on the modeest package. Now use a built-in function to estimate the mode of a posterior distribution.
  • Rewrite the "printFit" function.

hBayesDM 0.3.0

  • Made several changes following the guidelines for R packages providing interfaces to Stan.
  • Stan models are precompiled and models will run immediately when called.
  • The default number of chains is set to 4.
  • The default value of adapt_delta is set to 0.95 to reduce the potential for divergences.
  • The “printFit” function uses LOOIC by default. Users can select WAIC or both (LOOIC & WAIC) if needed.

hBayesDM 0.2.3.3

  • Add help files
  • Add a function for checking Rhat values (rhat).
  • Change a link to its tutorial website

hBayesDM 0.2.3.2

  • Use wide normal distributions for unbounded parameters (gng_* models).
  • Automatic removal of rows (trials) containing NAs.

hBayesDM 0.2.3.1

  • Add a function for plotting individual parameters (plotInd)

hBayesDM 0.2.3

  • Add a new task: the Ultimatum Game
  • Add new models for the Probabilistic Reversal Learning and Risk Aversion tasks
  • ‘bandit2arm’ -> change its name to ‘bandit2arm_delta’. Now all model names are in the same format (i.e., TASK_MODEL).
  • Users can extract model-based regressors from gng_m* models
  • Include the option of customizing control parameters (adapt_delta, max_treedepth, stepsize)
  • ‘plotHDI’ function -> add ‘fontSize’ argument & change the color of histogram

hBayesDM 0.2.1

Bug fixes

  • All models: Fix errors when indPars=“mode”
  • ra_prospect model: Add description for column names of a data (*.txt) file

Change

  • Change standard deviations of ‘b’ and ‘pi’ priors in gng_* models

hBayesDM 0.2.0

Initially released.