Hypothetical Knowledge and Games with Perfect Information

A standard model for a game with complete information consists of a state space with partitions, and a specification of the strategies played in each state. We show that such models are inadequate for explaining players' behavior. We propose instead extended models in which it is possible to express not only knowledge but also hypothetical knowledge, i.e., theories regarding subgames that are known not to be reached. In such models strategies are no longer primitives. Each state specifies behavior rather than strategies, and the latter are derived using hypothetical knowledge. In extended models common knowledge of rationality does not imply backward induction. We describe an intuitive condition that guarantees backward induction. Moreover, it is possible to express formally the idea that the theories that support the backward induction path assume some irrationality off path.