Some Computational Desiderata for Recognizing and Reasoning About the Intentions of Others∗

Reasoning about intentional action is a pervasive and critical skill in the human cognitive repertoire. In-tentions have taken center-stage in discussions of how humans parse perceptual input, understand language, make moral judgments, and predict the behavior of conspecifics. In the quest to engineer machine intelli-gence, intentions have largely either been ignored en-tirely, or have been given oversimplified construals as either preference-orders over actions, or as simple pred-icates in computational theories of action. In this paper, we will motivate the need for intelligent systems capa-ble of reasoning about the intentions of others by pre-senting a number of germane application areas, includ-ing those which deal with the integration of intention-recognition with other cognitive processes including di-alogue processing. We then briefly review the relevant psychological literature on the development and oper-ation of the human capacity to recognize and reason about the intentions of others. In doing so, we will extract a number of desiderata for the development of computational models of intention-recognition. We will then show how these requirements motivate principled design choices for the construction of intelligent sys-tems. Finally, we will close with a brief description of Polyscheme, a computational cognitive architecture which we feel sufficiently addresses the computational challenges entailed by our desiderata