A decision-theoretic model of coordination and communication in autonomous systems.

Abstract

This thesis presents a formal model of rational, autonomous behavior in single- and multi-agent domains, together with its implementation--the Rational Reasoning System (RRS). The model is based on a notion of economic rationality, which directs an agent to maximize the expected utility of its interactions with the environment. Rationality is implemented by having an agent repeatedly calculate the expected utility of plans of action available to it, and execute the plan with the highest utility. In cases in which the agent is interacting with other agents, the calculation of the expected utility includes the anticipated actions of others. An agent predicts the actions of the other agents using the assumption that they also are rational. The fact that they may be using the same assumption about the original agent leads to a recursive nesting of beliefs. The formalism that expresses this recursion is called the Recursive Modeling Method (RMM), which can be shown to converge on the best choice of action for an agent after considering a finite number of levels of recursive models. RMM can be used to compute the expected utilities of possible messages the agent can exchange with others, thus allowing the agent to be rational in its communicative behavior. This method can be extended to account for the possibility of the messages being dishonest and not believed. RMM has led to a preliminary version of a recursive variant of the possible world semantics that provides for a natural distinction between the concepts of knowledge and belief, and facilitates both recursively nested deductive reasoning, useful for problems like the Three Wise Men puzzle, and recursive decision-theoretic reasoning for coordination and communication.