Box 7—figure 1. An example of successful and unsuccessful model validation.

(A) Behavior is simulated by one of two reinforcement learning models (a blind agent and a state-based agent) performing the same learning task. Generative parameters of the two models were set so that the learning curves of the models were approximately equal. (B) Likelihood per trial seems to indicate a worse fit for the state-based-simulated data than the blind-simulated data. (C) However, validation by model simulations with fit parameters shows that the state-based model captures the data from the state-based agent (compare dark learning curves in panels A and C), but not from the the blind agent (yellow learning curves in panels A and C).