Lessons Learned: Generalizing Learning Across Games

Abstract
This paper synthesizes findings from an ongoing research program on learning in signaling games. The present paper focuses on cross-game learning - the ability of subjects to take what has been learned in one game and generalize it to related games - an issue that has been ignored in most of the learning literature. We begin by laying out the basic experimental design and recapitulating early results characterizing the learning process. We then report results from an initial experiment in which we find a surprising degree of positive cross-game learning, contrary to the predictions of commonly employed learning models and to the findings of cognitive psychologists. We next explore two features of the environment that help to explain when and why this positive transfer occurs. First, we examine the effects of abstract versus meaningful context, an issue that has been largely ignored by economists out of the belief that behavior is largely dictated by the deep mathematical structure of a game. In contrast, results from cognitive psychology suggest that behavior may well be sensitive to context employed. Our results show that the use of meaningful context serves as a catalyst for positive transfer. Second, we explore how play by two-person teams differs from play by individuals. The psychology literature is quite pessimistic about the ability of teams to beat a "truth wins" standard based on performance of individuals. But teams easily surpass this norm in our cross-game experiment. We use the dialogues between team members to gain insight into how this transfer occurs, gaining direct confirmation for hypotheses generated by econometric analysis of earlier data. I. The Experimental Environment: Our experiments are based on a simplified version of Paul Milgrom and John Roberts' (1982) entry limit pricing game. The game proceeds as follows: (1) Monopolists (Ms) observe their cost level - high (MH) or low (ML) cost - realized according to equal probabilities that are common knowledge. (2) Ms choose a quantity (output) whose payoff is contingent on the entrant's (Es) response (see Table 1). (3) E sees this output, but not M's type, and either enters or stays out. The asymmetric information, in conjunction with the fact that it is profitable to enter against MHs, but not against MLs, provides an incentive for strategic play (limit pricing). (Insert Table 1 here) As a treatment variable, Es use either the high or low cost payoff table. With high cost Es, there exist