Dimension-wise Sequential Update for Learning a Multidimensional Environment in Humans

When confronted with multidimensional environment problems, humans may need to jointly update multiple state-action-outcome associations across various dimensions. Computational modeling of human behavior and neural activities suggests that such updates are implemented based upon Bayesian update principle. However, it is unclear whether humans perform these updates individually or sequentially. If the update occurs sequentially, the order in which the associations are updated matters and can influence the updated results. To address this question, we tested a few computational models with different update orders using both human behavior and EEG data. Our results indicated that a model undertaking dimension-wise sequential updates was the best fit to human behavior. In this model, ordering the dimensions was decided using entropy, which indexed the uncertainty of associations. Simultaneously collected EEG data revealed evoked potentials that were correlated to the proposed timing of this model. These findings provide new insights into the temporal processes underlying Bayesian update in multidimensional environments.