Fig. 2
Results from an initial user study. In this study, participants were paired with other people, S++, and MBRL-1 in four different repeated normal-form games. Each game consisted of 50+ rounds. a The proportion of rounds in which both players cooperated when a player was paired with a partner of like type (self-play) and with a human. b The standardized payoff, computed as the standardized z-score, obtained by each algorithm when paired with each partner type. In both plots, error bars show the standard error of the mean. These plots show that while S++ learns to cooperate with a copy of itself, it fails to consistently forge cooperative relationships with people. There were some variations across games. Details about the user study and results are provided in Supplementary Note 5
