Fig. 2.

Model DIST fit to human performances. a Proportions of choosing true best bandits (maximising reward frequencies x proposed rewards) following reversals. b Proportions of choosing proposed rewards. In a and b, mean proportions are shown in the congruent, neutral and incongruent condition for fitted model DIST (green) and participants (blue, ±s.e.m.). See Supplementary Fig. 1 for (parameterised) model OPT fit. c Appetitive values of proposed monetary rewards $Ṽ$(left) and subjective probabilities in inferred bandits’ reward probabilities $\tilde{P}$ (right) in model DIST fitting human choices (lines correspond to distortion functions with mean parameters over participants). d Factorial analysis of choice log-odds over trials sorted according to subjective probabilities $\left({\tilde{P}}_1>{\tilde{P}}_2\mathrm{vs}\mathrm{.}{\tilde{P}}_1<{\tilde{P}}_2\right)$, appetitive values of proposed rewards $\left({Ṽ}_1>{Ṽ}_2\mathrm{vs}\mathrm{.}{Ṽ}_1<{Ṽ}_2\right)$ and total appetitive values (${Ṽ}_1+{Ṽ}_2$, median split) for fitted model DIST simulations (green) and human data (blue). Both model DIST and participants exhibited main effects of subjective probabilities and relative appetitive values (all Fs(1,21) > 23.7, ps < 0.00001) with no main effects of total appetitive values (both Fs < 1). However, model DIST unlike participants exhibited an interaction between subjective probabilities and total appetitive values (DIST: F(1,21) = 2269, p < 0.00001. participants: F < 1). e Same factorial analysis of choice log-odds adjusted for RL-values for fitted model DIST+RL and human data (see text). Model DIST+RL exhibited main effects of subjective probabilities, relative appetitive values and total appetitive values (all Fs(1,21) > 7.8, ps < 0.01), along with an interaction between subjective probabilities and total appetitive values (F(1,21) = 691, p < 0.00001). By contrast, participants exhibited only a main effect of relative appetitive values (F(1,21) = 18.5, p < 0.0001): participants exhibited no main effects of subjective probabilities (F(1,21) = 3.0, p = 0.10), and neither main nor interaction effects associated with total appetitive values (both Fs(1,21) < 1.2, ps > 0.29). See Supplementary Table 1 for model best-fitting parameters and Supplementary Fig. 2 for model-free analysis. Error bars are s.e.m. over participants (N = 22)