LANSKY@SRI-WARBUCKS.ARPA.UUCP (07/31/86)
DECISION-MAKING AND ACTION IN THE REAL WORLD
John Myers (JMYERS@SRI-AI)
SRI International, Robotics Laboratory
11:00 AM, MONDAY, Aug. 4
SRI International, Building E, Room EK228
In this philosophical talk I will present my opinions as to how to
design an entity capable of operating in the real world, under limited
resources. These include limited time, information, and capabilities.
I will present models that stress heuristic aspects of behavior,
rather than traditional pre-planning techniques. As Terry Winograd has
said, "The main problem is to come up with what you are going to do in
the next five seconds."
After covering the problem and some traditional paradigms, I will
discuss three main concepts, along with a follow-up concept. These
are: the Theory of Stances, the Freudian Motivation Model, and the
Theory of Alternative Choices, along with the Principle of
Responsibility. These are contrasted against traditional approaches
by their emphasis on workability, as opposed to correctness.
A Stance consists of a high-level classification of a situation, along
with a high-level precompiled response script. Often there is
insufficient information in a prima facia situation to correctly
determine what is going on; or, the entity may simply not be able to
afford the overhead required to completely plan its behavior from
first principles. Taking a stance on the situation allows a habitual
response to be made; which at least is some action in the face of the
unknown, and at best, solves the problem with minimal effort.
The Freudian Motivation Model splits behavior generation into three
general processes: generation, policies, and judgment, corresponding
to the id, superego, and ego, respectively. Approved behaviors are
put on an intention queue or a performance queue, among others. The
model can be used to explain nonpurposeful or nonvolitional behaviors
such as posthypnotic acts or compulsions.
The Theory of Alternative Choices says that given a direct choice
between, for example, one of two actions, there are actually a number
of alternative decisions that must be considered. These include: do
nothing, wait, waffle, observe/consult, relegate, delegate, react,
transcend, or respond with a stance. One of these may be much more
appropriate in a resource-limited situation than directly planning out
a decision between the two original choices.
As a follow-up, the Principle of Responsibility says that the entity
(the computer) must be responsible for its actions and its
recommendations. In a certain sense, it must be willing to be wrong.
Even if it is totally convinced of the correctness of its situational
assessment, it must consider the possibility that things might go
badly, given a certain course of action--and it must use that as
further input to the decision process.
Examples will be interspersed in the talk.
VISITORS: Please arrive 5 minutes early so that you can be escorted up
from the E-building receptionist's desk. Thanks!