A model is proposed to explain the results of recent experiments in which subjects repeatedly played a coordination game, with the right to play auctioned each period in a larger group. Subjects bid the market-clearing price to a level recoverable only in the efficient equilibrium and then converged to that equilibrium, although subjects playing the game without auctions converged to inefficient equilibria. The efficiency-enhancing effect of auctions is reminiscent of forward induction, but is not explained by equilibrium refinements. The model explains it by showing how strategic uncertainty interacts with history-dependent learning dynamics to determine equilibrium selection.