Separate learning rates showed no clear gain for later choices in ds004295 v1.0.0 with choice repetition, initial values of 0.5, and a 2 percent lapse

We tested one learning rate against two feedback-specific rates in reward-gain and punishment-avoidance reversal tasks from OpenNeuro ds004295 v1.0.0. The sample included 23 people with complete behavior arrays. Each model was fit on trials 1–140 and scored on trials 141–280, with choice repetition, initial values of 0.5, and a fixed 2 percent lapse. The mean test log-loss difference (one rate minus two rates) was −0.004876 natural log units per choice; the 95 percent paired person-bootstrap interval was −0.014675 to 0.002979. The interval includes zero, so this test does not establish a gain for two rates or prove equal performance. A planned check without choice repetition favored two rates, with a mean difference of 0.018588 and interval 0.006275 to 0.032479. The secondary intervals were not corrected for multiple tests. The observed result therefore depends on the tested choice rule. This is a new model comparison on an existing sample, not an independent sample. It makes no claim about EEG signals or brain mechanisms.

reversal learning · learning rates · choice prediction · log loss · reward · punishment avoidance · model comparison · OpenNeuro ds004295