Fig. 7. Bayesian model comparison.

a Model comparison between different family of models against optimal model (Bayes-DDM) using Bayes factor. Although the models were only fitted on the mean psycho- and chronometric function, the quality of the fitted data were evaluated considering an objective function additionally takes each model’s prediction of the conditional psychometric functions regarding previous reward into account (see Methods for details). Here we depict: two optimal DDM that do not learn weights, one with trial-by-trial bias modulation and one without; a random weights model in which stimulus-to-bound weights fluctuate from trial to trial stochastically; two simplified versions of Bayes-DDM: one without bias modulation and another without weight learning; and two heuristic models in which we implement a network that mimics a DDM that learns the stimulus-to-choice map through Reinforcement learning (RL-DDM), being one of them without confidence weighted learning. The gray dots show the Bayes factor for individual rats, and the black dots the Bayes factors for the fits using the accumulated data across all rats (adjusted for the increased number of trials). For more details about these models and other versions see Methods and Supplementary material.