Common Knowledge and the Theory of Interaction
Dissertation, University of Pittsburgh (
1988)
Copy
BIBTEX
Abstract
The dissertation examines the concept and current theory of common knowledge, with special emphasis on its significance for interaction. I begin by examining the rational choice theory of interaction more broadly, arguing for specific desiderata of the model. I then discuss the current rational choice model of interaction, which is non-cooperative game theory, and argue that common knowledge is an essential assumption on which game theoretic explanations of interaction hinge. Game theory's account of common knowledge, which stems from Lewis' Convention, is that common knowledge among rational agents involves a potentially infinite series of implicitly held mutual beliefs: we all know x, we all know that we know x, we all know that we know that we know x,$\...$ This account I call 'common knowledge*.' ;I advance two main objections to common knowledge*. First, common knowledge* cannot be 'bootstrapped.' That is, it always takes common knowledge* to get further common knowledge*. Second, I show that this account of common knowledge relies on the assumption, which I argue is untenable, that rationality entails common knowledge* of the possible states of the world. ;The theory of common knowledge developed by Lewis and the game theorists is attractive because it allows us to explain conventions and solutions to noncooperative games as individually rational. Furthermore, this model of interaction is a deductive model; one can deduce the rational course of action given the required inputs. However, without the common knowledge* assumption, the model can no longer be deductive. I also argue that a type of frame problem inevitably arises in a deductive model of interaction. I sketch an alternative, commonsense theory of common knowledge, which takes conventional behavior as primitive in explaining common knowledge, thus turning Lewis' view on its head. I tell a story to illustrate how the commonsense theory can be bootstrapped. I then suggest a program for building an inductive model of rational interaction