Data

Model Calibration (prior to the competition)

To help participants calibrate their models, we provided the behavioral data of human subjects participating in the competition's experimental task. 

• The dataset can be found in the competition's repository. 

• The dataset includes more than 500 participants, experiencing the following reward schedules:

• 100 subjects were tested in a bandit task, in which a reward was assigned to alternative 1 in 33/100 trials (randomly selected) and to alternative 2 in 17/100 of the trials. (randomly selected). Note that this data is useful for quantifying learning but the reward schedule does not comply with the rules of the competition.  

• 400 subjects were tested in schedules that comply with the constraints, namely with exactly 25 rewards assigned to each alternative. In order to allow participants to estimate the variability between subjects, we used 20 different random schedules, and each schedule was tested on 20 different subjects. 

Competition dataset 

• The full dataset of the competition can be found here. 

• Implementation of behavioral models, and schedule optimization procedures used in the competition can be found here.