Table 1:
Results of model comparisons using Bayesian information criterion
| Three-parameter model (σm, σe, rewarded, σe, unrewarded) | Two-parameter model (σm, σe) | One-parameter model (σe) | |
|---|---|---|---|
| Experiment 1 | 0 | 1735 | 392 |
| Experiment 2 | 0 | 389 | 272 |
| Cerebellar | 184 | 0 | 1232 |
Our model comprised three parameters: the SDs of the Gaussian distributions of motor noise (σm) and exploration following rewarded (σe, rewarded) and unrewarded (σe, unrewarded) trials. To examine the relative importance of each model parameter, we compared the full model to two reduced models: one where exploration variability does not depend on reward history (two-parameter model: σm and σe) and one that does not include motor noise (one-parameter model: σe). Model comparisons using BIC show the three-parameter model best fit the data from experiments 1 and 2, and the two-parameter model best fit data from the group with cerebellar damage. For each experiment, we show the difference in BIC relative to the best model (i.e., the one with 0).