Report Number: CS-TN-94-6
Institution: Stanford University, Department of Computer Science
Title: Ascribing Beliefs
Author: Brafman, Ronen I.
Author: Tennenholtz, Moshe
Date: December 1993
Abstract: Models of agents that employ formal notions of mental states are useful and often easier to construct than models at the symbol (e.g., programming language) or physical (e.g., mechanical) level. In order to enjoy these benefits, we must supply a coherent picture of mental-level models, that is, a description of the various components of the mental level, their dynamics and their inter-relations. However, these abstractions provide weak modelling tools unless (1) they are grounded in more concrete notions; and (2) we can show when it is appropriate to use them. In this paper we propose a model that grounds the mental state of the agent in its actions. We then characterize a class of {\em goal-seeking\/} agents that can be modelled as having beliefs. This paper emphasizes the task of belief ascription. On one level this is the practical task of deducing an agent's beliefs, and we look at assumptions that can help constrain the set of beliefs an agent can be ascribed, showing cases in which, under these assumptions, this set is unique. We also investigate the computational complexity of this task, characterizing a class of agents to whom belief ascription is tractable. But on a deeper level, our model of belief ascription supplies concrete semantics to beliefs, one that is grounded in an observable notion -- action.