, 2009, Sutton and Barto, 1998 and Watkins and Dayan, 1992) to each subject’s sequence of choices. To account for the observed decrease in learning, we implemented an exponentially decreasing half-life time as a free model parameter

that reduces the learning rate in later trials providing single-trial estimates of the learning rate (αt). Maximum likelihood estimated (MLE) learning parameters of the model did not differ for learning from real and fictive outcomes ( Table 1), indicating Crenolanib ic50 that subjects could utilize both sources of information with similar efficiency. This is also supported by the fact that sensitivity to misleading probabilistic feedback did not differ significantly between real and fictive conditions ( Supplemental Information available online). MLEs of the half-life time indicated an average decrease of αt of more than 90% in both conditions per block. Additionally, negative log-likelihood (−LL) did not differ when compared between good and bad stimuli. Submitting feedback-locked EEG epochs to multiple robust regression analysis (Cohen and Cavanagh, 2011, O’Leary, 1990 and Rousselet Trichostatin A et al., 2008) revealed a double

dissociation of cortical PE correlates between real and fictive outcomes in the first 400 ms following feedback. Intriguingly, the first significant covariation of feedback-locked EEG activity with PEs was found exclusively for fictive outcomes: a negative early occipital effect occurred 192–238 ms after feedback (Figure 2A and Movie S1) and was localized to extrastriate visual and posteromedial cortex (PMC; Figure S2A). Phosphoprotein phosphatase In contrast, only real outcomes were associated with a somewhat later positive early PE effect spanning from 236–294 ms and a subsequent negative midlatency frontal PE covariation at 336–430 ms, which in the averaged

event-related potentials (ERPs) give rise to the feedback-related negativity (FRN) and P3a components, respectively (Figure 3). Direct contrasts between both conditions showed significant differences at electrode Oz during the time window of the occipital PE effect (peak t30 = −4.18, 204 ms, p < 0.0005) and at electrode FCz during FRN (peak t30 = 4.95, 284 ms, p < 10−4), as well as P3a time windows (peak t30 = −7.95, 394 ms, p < 10−8) ( Figure 4B). The temporospatial double dissociation in early processing of real and fictive feedback was statistically confirmed by a triple interaction of the factors electrode, time window, and condition in an ANOVA on the average regression weights of the early PE effects in significant time windows (190–240 ms and 250–300 ms, for fictive and real feedback, respectively) at the most significant electrodes (Oz and FCz, for fictive and real feedback, respectively). The FRN is usually found on unfavorable outcomes that violate expectancies (Gehring and Willoughby, 2002 and Miltner et al., 1997).