Formal game theory rests on an implicit model of how players will behave in games. It is assumed that all players are very rational, using sophisticated reasoning processes about the game and opposing players’ strategies to reach a plan of action. Pushing this model of rationality ever farther, game theorists have reached strong conclusions about how players should behave in a variety of situations.

Signaling game refinements are a notable example of this phenomenon. Virtually all interesting signaling games have a plethora of sequential equilibria, making it nearly impossible to make meaningful predictions. Drawing heavily on the logic of forward induction, theorists have derived a number of equilibrium refinements for signaling games which reduce the set of admissible equilibria to a workable size. Unfortunately, most of these equilibrium refinements perform poorly in laboratory experiments (Brandts and Holt, 1992). Instead the evidence suggests that subjects are best modeled as learning to behave strategically through a gradual learning process (Cooper, Garvin, and Kagel, 1997). Thus, a good model of how signaling games (and games in general) are actually played must be based on solid empirical evidence of how individuals learn to play strategically.

One of the important questions raised by learning models as a vehicle for generating equilibrium outcomes is the ability of subjects to transfer their learning across related. The following quote from Fudenberg and Kreps (1988) characterizes the importance of the issue ". . . it seems unreasonable to expect the exact same game to be repeated over and over; put another way, if we could only justify the use of Nash analysis in such situations, we would not have provided much reason to have faith in the widespread applications that are found in the literature. Faith can be greater if, as seems reasonable, players infer about how their opponents will act in one situation from how opponents acted in other, similar situations." To this end, our current work explores the issue of learning generalizability in signaling game experiments.

In this paper we present a series of experiments in which subjects gain experience in one signaling game and are then switched to a second related signaling game. We are interested in how subjects’ experiences in the first game affect their ability to play strategically in the second game. Our conclusions are as follows: 1) Even for extremely similar games, the learning process is significantly slowed by switching between games. 2) The context in which games are played has a significant impact on transfer even though it is irrelevant for the underlying game. 3) While transfer is imperfect, subjects display an ability to generalize their learning about strategic play which belies any model of players as simple minded automata. 4) While experienced subjects demonstrate impressive strategic sophistication, they still aren’t perfect game theorists. As such, we are able to generate violations of all major equilibrium refinements for signaling games with experienced subjects.

Co-author John H. Kagel