A computational investigation of being in the world

Abstract

Is there a rational explanation for human behavior? Or is it fundamentally idiosyncratic and beyond the ability of science to accurately predict? In everyday life, we are able to predict the preferences of other people relatively well, and function in a society that is strongly predicated on our ability to do so. Theoretical efforts at predicting how people form preferences, however, have met with repeated failures, resulting in widespread pessimism regarding the possibility of a universal rational explanation for human behavior. In this thesis, we provide precisely such an explanation. We show that the errors plaguing existing systems of preference representations are a direct result of the mystery surrounding the actual act of {\em formation} of preferences, and that once this latter mechanism is clarified, a very large number of paradoxical and contradictory empirical results from the behavioral economics literature are theoretically reconciled. Our investigations lead us to believe that a combination of two simple natural principles is sufficient to both predict and explain why humans make the choices they do: one, that humans seek to always learn what to do in the most statistically efficient manner possible, and two, that this quest for understanding is constrained in remarkably systematic ways by a competing search for choices that can be made with minimal cognitive effort. We find, therefore, that the rational goal that best describes human choice behavior is attempting to minimize the cognitive effort required to make a decision. In other words, in this dissertation, we propose a theory that rational human action is governed by a universal explanatory principle, one that does not match traditional expectations of utility maximization - the principle of least cognitive effort. This redefinition of rationality has far-reaching implications. In order to better understand them, we constructed an information-theoretic description of a meta-cognitive agent engaging with its environment which allowed us to formulate computationally tractable intrinsically motivated agents. In this dissertation, we report simulation studies which confirm that the behavior of cognitively efficient agents provides a unified explanation for a large number of behavioral biases identified by behavioral and experimental economists in human subjects, as well as a number of variations in subjects' perception of risk observed in neuroeconomics studies.We further show using mathematical arguments, that this construction is in consonance with existing reinforcement learning literature and, in fact, subsumes multiple strands of current research in broadening the definitions of reward in reinforcement learning. Finally, we extend our analysis to studying social behavior among populations of agents and shed new light on paradoxes in game theory and theories of social interaction, resulting in a demonstration of an amoral basis for being good - the existence of an entirely self-interested (and non-evolutionary) basis for cooperative and altruistic behavior. In short, this thesis proposes a quantifiable description of agents {\em being in the world} - detailing universal principles that explain how and why beings develop preferences of the form they do given the structure of the world they inhabit. Our results provide a unification of explanations for several biological and behavioral phenomena spanning economics, psychology, neurobiology, cognitive science, artificial intelligence and metaphysics.