Table 3.

Model comparison results with only the five best models shown here. Each reinforcement learning model had a single learning rate and inverse temperature parameter. Added to this base model was perseverance, effector bias, separate learning rate for positive and negative feedback (‘neg α’), a lapse rate, and exponential decay for unchosen options back to $Q=0.5$. The integrated Bayesian Information Criterion was estimated for 200 k samples each from the practice and scanning session, and summed over both sessions and participants to arrive at final BICi.

Additional parameters	BICi	δBICi
neg α, decay	12393	0
perseverance, neg α, decay	12400	+7
lapse rate, neg α, decay	12427	+34
perseverance, lapse rate, neg α, decay	12435	+42