Full 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 tested choice forecasts in 23 people from OpenNeuro ds004295 v1.0.0. Each person completed two 280-trial reversal tasks. We fitted the first 140 trials and scored the next 140 trials. Both learning models had one learning rate, a value coefficient, and a repetition coefficient. One model updated only the chosen option. The other also updated the unchosen option toward the opposite outcome. Both started at 0.5 and used a fixed 2 percent lapse. The mean log-loss gain from the opposite-outcome rule was 0.013825 nats per later choice. The person bootstrap 95 percent interval was [-0.001044, 0.029969]. This interval included zero. A simple last-outcome rule had a lower mean test loss than either fitted learning rule. Some planned settings gave intervals above zero. The main result does not establish equal performance or a human learning mechanism. We also prove the last-outcome limit of full opposite-outcome updating. The package contains source data, code, forecasts, fits, checks, and retrieval instructions.

reversal learning · choice forecasting · reinforcement learning · opposite-outcome updating · repetition · conditional prediction · log loss · person bootstrap