Cognition as a sequential decision problem

Abstract

How should we attempt to understand the mind? Historically, there have been two broad approaches. The \emph{rational} approach focuses on characterizing the problems people have to solve and the optimal solutions to those problems, explaining \emph{why} people behave in the way they do. In contrast, the \emph{mechanistic} approach focuses on identifying the cognitive processes underlying behavior, explaining \emph{how} the mind actually works. Traditionally, these approaches have been viewed as conflicting, but recent years have seen a growing interest in models that synthesize the two approaches. This dissertation presents a formal framework for deriving models of cognition that are both rational and mechanistic. The key idea to broaden the concept of the ``environment'' to which cognition adapts: cognitive processes are adapted not only to the external environment (the world), but also to the internal environment (the brain). Formalizing this old idea, I cast cognition as a sequential decision problem in which an agent executes cognitive actions to navigate between mental states and, ultimately, produce effective behavior. In three domains---attention, memory, and planning---I show how the framework can be applied to yield models that explain both how the mind works and why it works that way.