hBayesDM 1.2.1
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.