3, p = 0.05 corrected; Table S3). This further shows that evidence-based aPEs are related to subjects’ behavior. We constructed a weighted semi-Bayesian variant of our sequential model to assess to what extent subject behavior was influenced by the evidence-based update as compared to the simulation-based update. This model included two additional free parameters, ρ and σ, that denote, respectively, the weight given to the simulation-based and evidence-based updates. See Supplemental Information for details. These parameters were estimated for each subject, and they effectively shift the distributions

on ability up or down relative to the Bayesian sequential model (Figure S6). To compute a between-subject covariate that reflected

the relative weighting Fulvestrant of the evidence-based update, we normalized the relevant Epigenetics inhibitor term by the sum of the two: σ/(ρ+σ). We found an overlapping region of rdlPFC that exhibited a strong relationship between this behavioral index and evidence-based aPEs (Figure 6B; Z = 2.3, p = 0.05 whole-brain corrected; Table S3). Moreover, analysis of independently identified ROIs revealed that this between-subject correlation was evident for both people (r = 0.58; p < 0.005) and algorithms (r = 0.48; p = 0.01). These analyses demonstrate that activity in the rdlPFC region correlates better with evidence-based aPEs in those individuals whose behavior is influenced more heavily by the evidence-based update than by the simulation-based update, further linking the neural signals and learning behavior. Agent performance can be attributed Ketanserin to ability or to chance. The behavioral regression analyses reported above show that subjects differentially credited specific agents for their correct and incorrect predictions in a manner that depended on the subjects’ own beliefs about the state

of the asset. We investigated the neural processes associated with this effect, by searching across the whole brain for regions exhibiting significant effects of the following contrast between unsigned aPEs at feedback: ((AC−DC) − (AI−DI)) × people − ((AC−DC) − (AI−DI)) × algorithms. Significant whole-brain corrected clusters were found in left lOFC and mPFC only (Figure 7; Z = 2.3, p = 0.05, corrected; Table S3). Importantly, this analysis controls for differential updating between people and algorithms that is simply due to (1) correct versus incorrect predictions (because DC trials are subtracted from AC trials), and (2) predictions with which subjects would likely agree versus disagree (because AI−DI trials are subtracted from AC−DC trials). Moreover, there was a strong between-subject correlation between the behavioral interaction effect illustrated in Figure 2D and the neural interaction effect in independently defined lOFC ROIs (r = 0.55; p < 0.01).