Abstract
This paper studies evolution of peoples’ models of how other people think - their theories of mind. For the case of games that are played for the first time, people are assumed to form beliefs according to the level-k model. This model postulates a hierarchy of types, such that an individual of type k plays a k times iterated best response to the uniform distribution. For the case of learning, it is assumed that the lowest type behaves in accordance with fictitious play, and that there is a hierarchy of more sophisticated types, which play iterated best responses to this. The models are also extended to allow for partial observability, in the sense that a higher type recognize and best respond to lower types, but not vice versa. Evolution according to the replicator dynamic is studied both across and within games. It is found that evolution may lead to stable states where different types, including low types, co-exist. This holds even when types are not observed.