Model comparison.

<p>Each bar shows the differential score relative to the model with the lowest score (log scale). Note that these scores are for the group as a whole. <b>Top</b>: Models 1–7 were fitted to the instrumental data only. Model 1 was a standard Rescorla-Wagner type model which forced rewards and punishments to be equally informative. It assumed equally fast learning about rewards and punishments, and no biases. Inclusion of either separate reward and punishment sensitivities (, Model 2) or separate biases in the approach and withdrawal blocks (Model 4) improved the fit. Separate learning rates for rewards and punishments (Model 3) did not improve the fit as much as separate reward and punishment sensitivities (Model 2). The best model (5) included a separate go bias in the approach and withdrawal blocks, and separate reward and punishment learning rates. Models that additionally allowed separate reinforcement sensitivities (Model 6), or separate learning rates (Model 7) in the approach and withdrawal blocks failed to improve the fit. <b>Bottom</b>: Comparison of models on both instrumental and PIT choice data jointly. Models 8–10 used the instrumental component of Model 5. Models 8–10 included ten Pavlovian factors, capturing the effect of each of the five Pavlovian stimuli in each of the two blocks. Model 9 allowed for extinction by including an exponential decay of the instrumental values during the PIT part of the task. Model 10 included random generalisation noise and provided the best fit.</p>




CC BY 4.0