Fig. 3.

(a) Mean predictive probabilities for all models. All models that include the ‘rho’ parameter fit the data better than the corresponding models that do not contain this parameter. (b) Model comparisons using each model's Bayesian Information Criterion (BIC_int). Despite the fact that the model that predicts the highest proportion of participants' choices is the ‘Pruning and Pavlovian’ model that contains the ‘rho’ parameter, this model is penalized due to its added complexity. The most parsimonious (i.e. ‘winning’) model is therefore the ‘Pruning’ model that includes the extra ‘rho’ parameter.