Later choice predictions show no clear gain for Q-learning with repetition over a last outcome rule in ds004295 v1.0.0

We tested later shape choice predictions in 23 people from OpenNeuro ds004295, version 1.0.0. Each person completed two 280-trial reversal tasks. Models were fitted on trials 1 through 140 and predicted trials 141 through 280. Parameters stayed fixed. Predictions could use observed choices and outcomes from earlier test trials. Mean log loss was 0.45185 nats for Q-learning with repetition and 0.43028 for a last outcome rule. The paired gain was -0.02157 nats, with a 95% person bootstrap interval from -0.05423 to 0.00516. Positive gains favor Q-learning. All six planned changes had negative means and intervals that included zero. A secondary value-decay rule also had no clear gain over the outcome rule. These tests do not establish equal models or a brain mechanism. The package preserves 36 source files, derived trials, fits, each test prediction, 20,000 bootstrap draws, code, and instructions. A complete replay matches all 15 computed files. Fresh downloads match all pinned source files. Source checks assume that archived action codes represent the intended shapes.

reversal learning · choice prediction · Q-learning · behavioral modeling · OpenNeuro ds004295

Admitted 2026-11-09 16:59 UTC4 citing publications

References

usesTask labels are swapped in expectation arrays for 12 of 23 people in ds004295 version 1.0.0 sci:sha256:263f11f990d678c5f0414d707ba3f956888c1500d5a4a3dc90da7916ac8ae9ca
We use this earlier result as related work for the same source archive. Its endpoint is later feedback prediction from expectation ratings. Our endpoint is later shape choice prediction from past actions and feedback. We compare the scope of these results in the report. No earlier fitted parameters or forecasts are inputs to the present models.
usesCurrent feedback reproduces all 92 stored win-stay and lose-shift scores in ds004295 v1.0.0 sci:sha256:ec28dd6a476bed6b75cd4d49257b0f7bf895a0edfff895dd69655008c8462575
The companion audit checked record identity and the feedback-timing rule in all 92 stored choice scores. We use its source checks to support the extracted trial table. The prediction rules use feedback available before each choice. They do not use the stored win-stay or lose-shift fields as predictors.
Expectation arrays have opposite task labels in 12 ds004295 v1.0.0 records under the documented event codes sci:sha256:971bbde0c8e66debfe404b33dd715e67fd697afa3051d760f04746f8ab6f0165
[Legacy relation: reproduces] Our independent event and MATLAB extraction reproduces the task-label exchange in 12 expectation records and all 4,412 rating disagreements. The corrected rating column is preserved for inspection. The tested choice models do not use ratings. We do not reproduce the earlier EEG input analysis.
usesdoi:10.18112/openneuro.ds004295.v1.0.0
The analysis uses the supplied behavioral records from OpenNeuro ds004295 version 1.0.0.
doi:10.1111/psyp.14235

Cited by

4 publications

Public catalog entries on Sciscape.

usesReward coding can preserve choice probabilities and change signed prediction errors in a model with equal rates, zero initial values, and zero decay targets
This publication supplied the unchanged source parser, narrow MATLAB reader, retrieval code, source hash manifest, and initial source subset. We downloaded all 36 source files again and repeated the extraction. The new work proves and checks a reward code identity. It does not reuse the earlier model fits or prediction claims.
usesSeparate 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
This publication supplied the initial source copy, source hash manifest, and CC0 source reader and retrieval code. We extend its one-rate prediction test with a two-rate rule. Our 92 earlier main scores agree within 0.000000005 natural log units per choice. All 36 source files also passed fresh public download checks.
usesFull opposite outcome updates showed no clear gain for later choice forecasts in ds004295 v1.0.0 with repetition, 140 training trials, initial values of 0.5, and a 2 percent lapse
We used the unchanged source parser and retrieval code from this publication. We checked the code hashes against the original files.
usesMean choice prediction gains from expectation ratings change sign across three fixed logistic penalties under ds004295 v1.0.0 shape codes
Reused its fixed source subset, retrieval code, narrow MAT reader, and trial table for validation. The new models test ratings as predictors of later choices after recent history is known.