🚀 go-pugleaf

           Department of Computer Science, University of Toronto
            (LP = D. L. Pratt Building, 6 King's College Road)

       -------------------------------------------------------------

                                    AI
                   LP266, at 10:00 a.m., 4 January 1993

                              Craig Boutilier
      Department of Computer Science, University of British Columbia

                          " What Should You Do? "

The ultimate aim of research in AI is to produce artificial agents that "do
the right thing" in whatever circumstances they happen to find themselves.
One crucial component of doing the right thing is believing the right
things (whether explicitly or implicitly); the second is "acting the right
way" given what you believe.  In this talk I will focus on qualitative
representations of an agent's beliefs and preferences, and the interaction
between the two.

To begin, I will discuss what an agent should believe, especially what
defeasible, or default, conclusions it should reach based on its certain
(for example, perceptual) knowledge. I will argue that conditional logics
provide a natural mechanism for the formal representation of default
beliefs, providing a qualitative analogue of probability theory.

I will then turn my attention to deontic logic, a branch of philosophical
logic that studies the notions of obligation and permission. We adopt the
view that any concept of preference on alternative outcomes can be thought
of as a value system that dictates how an agent ought to act. We simply
think of an agent's designer as specifying the appropriate "moral code" for
the agent. Certain puzzles in deontic logic have naturally lead to their
formalization as conditional logics. Somewhat surprisingly, the system I
will advocate is identical to that I propose for default reasoning.

Finally, I will present some preliminary results on how the two concepts
interact. I will demonstrate that preferences cannot alone dictate the
obligations (actions) of an agent. The belief state of an agent can
radically change its (desired) behavior.  This leads to the notion of a
conditional goal which is quite distinct from the type of goals typically
studied in AI. I will also show that once we take into account the
"abilities" of an agent (for an agent cannot be obligated to do something
it physically cannot!), the methods of deontic logic become under-
constrained. We are lead to various "strategies" for action.  I will show
that an agent can improve its behavior by making observations or increasing
its knowledge of the world. I will suggest the directions in which this
work can be extended, especially how we might investigate planning under
uncertainty within this formal framework.