MFIT

Simple model-fitting tools.

Questions? Contact Sam Gershman (gershman@fas.harvard.edu).

—————

Quick start. Examples below use snippets from mfit_demo_RL.m:

1. Define your prior by constructing a parameter structure. Here is an example from a reinforcement learning model:

g = [2 1];  % parameters of the gamma prior
param(1).name = 'inverse temperature';
param(1).logpdf = @(x) sum(log(gampdf(x,g(1),g(2))));  % log density function for prior
param(1).lb = 0;    % lower bound
param(1).ub = 50;   % upper bound

a = 1.2; b = 1.2;   % parameters of beta prior
param(2).name = 'learning rate';
param(2).logpdf = @(x) sum(log(betapdf(x,a,b)));
param(2).lb = 0;
param(2).ub = 1;

Here logpdf takes as input a parameter (or multiple parameters, in the case of multiple subjects) and evaluates the log joint density. The fields lb and ub correspond to the lower and upper parameter bounds, respectively.

2. Call the optimizer, which finds the maximum a posteriori estimates of the parameters for each subject:

results = mfit_optimize(@fun,param,data,nstarts)

Here @fun is a function handle for your log-likelihood function, which takes the following form:

lik = fun(x,data,options)

where x is a vector of parameter values, data is a single subject's data structure. It must have a field called N (i.e., data.N) that specifies the number of observations for the subject. options is an additional input structure that may be omitted.

See mfit_demo_RL.m for examples of Bayesian model comparison and cross-validation.