Decision-Theoretic Deliberation in Resource Bounded Self-Aware Agents

In this paper we introduce the notion of decisiontheoretic deliberation, which studies the relation between classical game and decision theory on the one hand, and agent theories of deliberation on the other hand. We aim at modelling the commonsense notion of intention in systems which are self-aware of their bounded reasoning power concerning their decisions. We propose a transparent decision theoretic deliberation model that contains besides standard actions that change the world and standard actions that only change the information state of the agent (capturing the standard notion of the value of information), also actions that change the agenda of the agent. The agenda contains book keeping of the agent’s decision theoretic planner, for which we use the DRIPS planner, and is used to deal with bounded resources, such as the trade-off between further deliberation or starting to execute actions, or whether to change the agenda in case of unexpected observations. This model is the basis of our study of intention. Creating an intention corresponds to putting something on the agenda, and reconsidering intentions corresponds to removing things from the agenda. We show that, in contrast to Cohen and Levesque’s approach that defines intention in terms of commitment, our treatment of commitment is based in terms of well understood concepts that consider both creation and reconsideration of intention. Consequently, intention cannot be defined locally in a state of the world, but only globally. Moreover, in contrast to abstract approaches such as Rao and Georgeff’s characterization of intention in terms of commitment strategies, we relate intention to planning. Finally, in contrast to Bratman’s theory of intention, we do not restrict ourselves to classical planning but we incorporate insights from more recent decision-theoretic planning. We illustrate our decision theoretic deliberation model and the related characterization of intention by a detailed example.