Recently there has been much theoretical and experimental work on learning in games. However, learning usually means "learning about the strategic behavior of opponents" rather than "learning about the game" as such. In contrast, here we report on an experiment designed to test whether players learn a) the relevant features of the payoff structure of a 2x2 game (e.g. the best reply structure) and b) whether they play according to equilibrium predictions regardless of how question a) is being answered. Question a) is being tested by asking subjects to guess the payoff structure of their opponents in a repeated encounter and rewarding subjects for correct answers.