Fig. 7: Steps in model construction procedure.

In the first step (1), models with one versus two learning rates were compared, and separately, models with an initial versus constant action initiation bias were compared. A fifth model was then constructed by combining all parameters from the winning models in step 1 (i.e., one versus two learning rates and the winning action initiation bias). In step 2, we tested whether model 5 was improved by adding a single magnitude sensitivity parameter (model 6) or separate magnitude sensitivity parameters for reward versus punishment outcomes (model 7). Finally, to confirm that the winning model from step 2 was the best overall model, we compared models 1–7 directly in step 3.