Luck and Learning in Games
Abstract: In strategic environments, actions and luck often combine to determine outcomes. We conduct a laboratory experiment with a version of the repeated O’Neill game in which (i) realized payoffs depend on both actions and luck, and (ii) players observe the effect of actions and luck separately. There is no reason for players to react to luck since it provides no information about either their opponent’s future actions or the future realizations of luck. We find, however, that subjects reinforce on luck. Specifically, a subject is more likely to choose the same action again if either it was ex-post optimal given his opponent’s action or, though suboptimal, resulted in a favorable payoff due to luck. These reinforcement effects are sub-additive. We estimate and compare several models of learning.
here.