A group of individuals repeatedly plays a fixed extensive-form game, u
sing past play to forecast future actions. Each (asymptotically) maxim
izes his own immediate expected payoff, believing that others' play co
rresponds to the historical frequencies of past play. Because players
observe only the path of play in each round, they may not learn how ot
hers act in parts of the game tree that are not reached infinitely oft
en. Hence, differences and correlations in beliefs about out-of-equili
brium actions may persist indefinitely. The stable points of these lea
rning processes are self-confirming equilibria, a weaker solution conc
ept than Nash equilibria. (C) 1995 Academic Press, Inc.