An R package (in progress and largely undocumented/untested, user beware) implementing the efficient simulation/estimation of reinforcement learning models of the Iowa Gambling Task (IGT), or n-armed bandit tasks more generally. The package implements several utility, updating, and temperature functions, and allows them to be combined arbitrarily to create new models. Both simulation and estimation use Stan as a backend, making the package very quick. The package can generate either full posterior distributions for parameters, or penalized ML estimates and CIs.
The package can be installed and loaded using the devtools package with
install_github(“areshenk/igt”, args = "--preclean”) library(igt)
Documentation and testing aren’t done yet, but the general operation looks like this
Decks are specified using the s4 class igt, which can be initialized using
x <- setupIgt()
by default, the function initializes a standard set of 4 decks, although custom configurations can be specified. We can get basic information about the task using
summary(x) > A B C D > EV -25.0000 -25.0000 25.00000 25.00000 > SD 135.3074 376.8892 27.52409 75.37784 > Win.Freq 0.5000 0.9000 0.60000 0.90000
numDecks(x) > 4
numTrials(x) > 100
Suppose that we want to simulate performance on the deck configuration using a model with a prospect utility function, and delta learning rule, and a trial independent temperature function. Then we can specify
utility <- 'PU' updating <- 'DEL' temperature <- 'TIC' pars <- list(utility = list(A = .25, l = .25), updating = list(r = .75), temperature = list(c = 1.2))
and simulate 100 players using
sim <- simulateIGT(x, n = 100, utility, updating, temperature, pars, scale = .01, full.output = FALSE)
scale = .01 indicates that wins and losses should be divided by 100 (as is standard), and
full.output = FALSE returns only the trial choices, wins, and losses (otherwise, the function will return the deck valuations, utilities, and choice probabilities for every trial).
sim is a list containing arrays
To fit a model to observed data, we first specify the utility, updating, and temperature functions (already done above), and create a list specifying the desired parameter bounds.
pars <- list(utility = list(A = c(0,1), l = c(0,5)), updating = list(r = c(0,1)), temperature = list(c = c(0,2))) fit <- estimateIGT(choice = sim$choice[1,], win = sim$wins[1,], loss = sim$losses[1,], n_decks = 4, utility, updating, temperature, pars, scale = .01, reg = 1, method = 'bayes')
reg is a regularization term. All parameters are initialized with support on the unit intervals and a
beta(reg,reg) prior. These raw parameters are then scaled to the width specified by the user. A value of
reg = 1 is equivalent to a uniform prior, with larger values concentrating more tightly in the center of the parameter space. RL models of the IGT have an annoying tendency to cluster towards the boundaries of the parameter space, owing largely to most models of the IGT being crappy models. I would generally recommend using something like
reg = 2 (at least) in practice, especially for ML estimation.
method = "bayes" causes the function to return a
stanfit object with the full posterior samples for all parameters. If point estimates are needed, then ideally the user would still just set
method = "bayes" and take the posterior means, but if speed is a concern, then
method = "map" returns maximum a posterior estimates along with (in the near future) confidence intervals. With the
reg term, these can be though of as l2 regularized ML estimated, with shrinkage towards the center of the parameter space.
Adding new functions
The package is designed to make it relatively simple to add new utility/updating/temperature functions. I plan to make a more complete guide in the future, but the basic steps for adding a new utility function are as follows (adding new updating/temperature functions works the same way, mutatis mutandis):
- Add the new utility function to
inst/chunks/utilityFunctions.stan. Note that all utility functions must have the same signature! Look at the other functions to see how this works
- Add a new conditional statement to the end
inst/chunks/utilityWrapper.stan. Again, the function is simple, so you'll see how it works.
- The package contains an internal lookup table in the form of a list called
modelComponentDetails, which contains information about the number of parameters in each utility function, bounds on those parameters (if any), and the parameter names. You'll need to add the necessary information to the list and save it to
R/sysdata.rda. Note that order is important: The order of functions in
modelComponentDetailsmust be identical to the order in
Then recompile/reinstall. Everything else should work more of less automatically.