fitr 
Python implementation of package to fit reinforcement learning models to behavioural data
Tutorial
Import package components
- Assumes your current directory is the
fitrfolder
Generate some synthetic data
import numpy as np
from rlparams import *
from tasks import bandit
nsubjects=50
# Initialize array of parameter values [learning rate, choice randomness]
params = np.zeros([nsubjects, 2])
# Sample parameter values from rlparams objects
params[:,0] = LearningRate().dist.rvs(size=nsubjects)
params[:,1] = ChoiceRandomness().dist.rvs(size=nsubjects)
# Run task
res = bandit().simulate(nsubjects=nsubjects,
ntrials=100,
params=params)
# Plot the cumulative reward
res.plot_cumreward()
# Scatterplots of total reward vs parameter values
res.cumreward_param_plot()
Fit a reinforcement learning model to the data
from fitr import *
from loglik_functions import bandit_ll
# Model with learning rate and choice randomness
lrcr_model = fitrmodel(loglik_func=bandit_ll().lr_cr,
params=[LearningRate(),ChoiceRandomness()])Now that the models are created, we can fit them. The default method (the only one presently implemented) is Expectation-Maximization.
lrcr_fit = lrcr_model.fit(data=res.data)Then you can plot the actual vs. estimated parameters as follows:
lrcr_fit.plot_ae(actual=res.params)Or you can also plot histograms of the parameter estimates:
lrcr_fit.param_hist()You can also plot the progression in Log-Model-Evidence, BIC, and AIC (whole model, not subject level) over the course of model fitting. LME should increase and then plateau, whereas BIC and AIC should decrease, then plateau. If there are deviations in the opposite direction for any of those, model fitting can be run again.
lrcr_fit.plot_fit_ts()
