Fig. 4: Reinforcement Learning models design and results.
From: Visual perceptual learning of feature conjunctions leverages non-linear mixed selectivity
a Mixed channel vs. Separate channel Reinforcement Learning models (MRL vs. SRL). In MRL, decision variables, DVot and DVct, are updated based on the reward prediction error computed over a joint Expected Value of the trial (EVt), while in SRL, the two channels are updated independently from each other based on their corresponding EVs (EVot or EVct). b, c MRL and SRL models’ performance compared with subjects’ performance on the test day of Experiment 1 and 2, respectively. d Models’ deviances from subjects’ choices on the Test day. MRL had smaller deviance as compared to SRL in Experiment 1 and vice versa in Experiment 2 (signed-rank test, MRL vs. SRL, Experiment 1: p = 0.0003, d = 2.0030, Experiment 2: p = 0.0038, d = −0.9100). MRL Mixed channel Reinforcement Learning, SRL Separate channel Reinforcement Learning, Sb: human subjects. * denotes p < 0.05, ** denotes p < 0.01, *** denotes p < 0.001, n.s. denotes non-significant. Error bars represent the standard error of the mean.
