Result of the model comparison procedure.

The trial-and-error learning task was performed by N = 85 subjects. For each subject, it was tested in descending order (see main text for details) which model provided the best fit for the initial learning phase. For 43 subjects (50.6%), the DRP models outperformed the FOP, BP and Q-learning models, with 36 subjects following the dfkl response pattern and 7 subjects following the lkfd pattern. Of the remaining subjects, 18 subjects (21.2%) showed a tendency towards generic optimal learning, while 19 subjects (22.3%) partially exploited stimulus-response dependencies. Q-learning was never significantly better than FOP or BP on the initial learning phase. Five subjects (5.9%) could not be assigned to a model-specific subsample.