< Back to Article

Nucleus accumbens dopamine release reflects Bayesian inference during instrumental learning

Fig 2

Bayesian and reinforcement learning models.

(A) Relationships between cognitive models (see Methods for more details). (B) Confusion matrix outlines results for model identification analysis. Each entry i, j represents the percentage of time that the column j fitting model best explained data generated by row i simulating model. The row orders are sorted via dendrogram based on model similarity (see Methods). (C) Model comparison using relative AIC compared to RL4p: , with lower values indicating better fit. (D) Illustration of value computation for BRL model family, which updates beliefs via Bayes’ rule and then uses these beliefs to compute values. (E) Illustration using a four-trial sequence (similar to [31]) to show the differences between RL4p and BRL. Top: purple and cyan bars show the choice values conditioned on the belief state; Bottom: pie charts show the belief state for BRL; the animal’s policy is selected as a function of the value within their belief states. (F) Behavior of different models compared to mouse data (black line). Trial 0 is when the program has switched the rewarded side in a block switch. (G) Example behavioral trajectory (probability of choosing the rightward port) predicted by different models. Mouse data are marked by a dashed line and block structure is marked by a solid line. Rewarded trials are marked as green dots and unrewarded trials are marked as red dots. Error bars show 95% bootstrapped confidence intervals.

doi: https://doi.org/10.1371/journal.pcbi.1013226.g002