With model values set to 0.5 at the start and a 2% lapse rule, two learning rates reduced new-pair choice loss in 77 ds004532 Study 1 adults
We tested two choice models in 77 adults from OpenNeuro ds004532 Study 1, version 1.2.0. Fits used only training choices and feedback. The test used 7,358 valid choices between new pairs of the same six stimuli. With all model values set to 0.5 at the start and a fixed 2% lapse rule, two learning rates reduced mean loss by 0.086629 nats per choice. The paired 95% person bootstrap interval was 0.049596 to 0.127702. The gain did not hold with starting values of 0 or 1. A reward-frequency rule gave loss close to the two-rate model. This result concerns prediction in one task. It does not identify a neural mechanism. The package preserves source files, code, every prediction, model fits, bootstrap draws, figures, and reproduction checks. A complete repeat from fresh source files matched all 13 numerical output files by hash.
probabilistic selection task · choice prediction · reinforcement learning · learning rates · model comparison · behavioral data · bootstrap
References
3- usesThe Choose A minus Avoid B score had opposite mean signs for C/E and D/F in 77 ds004532 Study 1 participants sci:sha256:54f30e78401a97533f0eab68074b546e76b1c4e84d203aef4ab436b29833e136
We used the 147 preserved source files, source manifest, and standard-library retrieval script from this report. These files enabled a new training-to-test model comparison. We downloaded each source file again and verified its hash. We did not use the earlier scores or analysis code.
- usesdoi:10.18112/openneuro.ds004532.v1.2.0
Study 1 training choices and outcomes were used to fit the models, and test choices from the same participants were used to score predictions. The archived source files match the pinned version and recorded Git tree.
- doi:10.7554/eLife.49547
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesThe cabergoline minus placebo score change interval included zero in 26 ds004532 adults under common additive visit effects without carryover
We used its checked source files and pinned source manifest. Its numeric event decoder informed our new decoder. We did not use its fitted models or effect estimates. We then retrieved all 64 required source files afresh and repeated the new analysis.
- usesA trained-pair term increased mean test loss under ridge logistic fits in 38 ds004532 Study 1 participants
We used this related report to distinguish training-to-test learning model prediction from the current test-phase choice model check. We read it after fixing our plan and results. Its scores did not select our sample or model rules. We did not compare losses across the two designs.