We report on experiments that tested the predictions of competing theories of learning in games. Experimental subjects played a version of the three-person matching-pennies game. The unique mixed-strategy Nash equilibrium of this game is locally unstable under naive Bayesian learning. Sophisticated Bayesian learning predicts that expectations will converge to Nash equilibrium if players observe the entire history of play. Neither theory requires payoffs to be common knowledge. We develop maximum-likelihood tests for the independence conditions implied by the mixed-strategy Nash equilibrium. We find that perfect monitoring was sufficient and complete payoff information was unnecessary for average play to be consistent with the equilibrium (as is predicted by sophisticated Bayesian learning). When subjects had imperfect monitoring and incomplete payoff information, average play was inconsistent with the equilibrium.