Dynamic Learning in a Two-Person Experimental Game

We analyze learning in a two-person simultaneous choice repeated game. Agents form beliefs about their rival's strategies, which they revise as the game progresses. We test this model using data from a series of experiments. We obtain data on subjects' expectations by asking them to predict their rival's choice in each period of the game; subjects are paid for each correct prediction. We analyze the data using the Kalman filtering technique, and find that learning is broadly consistent with Bayesian updating. The predicted equilibrium based on this learning process is indistinguishable from observed behavior for most subjects in later periods