Thursday, April 30, 2009
2:00 PM
Computer Studies Bldg. Room 632
Ph.D. Thesis Proposal
Daphne Liu
University of Rochester
Integrating Intrinsic Motivation into an Explicitly Self-Aware Dialogue Agent
Most work in AI intelligent agents lacks a well-balanced integration
of self-motivation into planning and reasoning. Planning and
reasoning have traditionally been viewed as being aimed at achieving
given user goals, without allowance for subjective evaluation of the
results of choices made. Conversely, most self-motivated agents
function according to a utility-maximizing policy mapping states of
world to the best actions to take in those states. However, the
learning of the policy and the policy itself do not involve
reasoning about actions and the expected action consequences.
Moreover, the states of these agents typically do not have a
symbolic representation amenable to general inference mechanisms.
We propose an explicitly self-aware and deliberately self-motivated
dialogue agent that integrates intrinsic motivation into its
planning and reasoning. Our agent is primarily driven by its desire
to optimize its own cumulative utility. It thinks ahead, plans and
reasons deliberately, and acts reflectively, by drawing on its
knowledge and by evaluating its action choices using a projective
lookahead and its own metrics of rewards/costs. Furthermore, our
agent can be seen as an integration of the behavioral (purely
opportunistic) agent paradigm and the planning-based (goal-directed)
agent paradigm.