SJG@SU-AI.ARPA (Matthew Ginsberg) (02/27/86)
From: Matthew Ginsberg <SJG@SU-AI.ARPA>
In light of what I expect will be department-wide interest in the
following talk, this week's research meeting/seminar of the KSL will
instead be a department-wide event.
The talk will run from 12.05 until 1.00 on February 28 and will be held
in the Chemistry Gazebo. The room is fairly small, so anyone interested
in attending would be well advised to arrive early.
Matt Ginsberg
FROM SOCRATES TO EXPERT SYSTEMS: THE LIMITS OF
CALCULATIVE RATIONALITY
BY
Hubert L. Dreyfus
University of California
Berkeley
An examination of the general epistemological assumptions behind
Artificial Intelligence research with special reference to recent
work in the development of expert systems. All AI work assumes that
knowledge must be represented in the mind as symbolic descriptions.
Expert system builders further assume that expertise consists in
problem-solving and that problem-solving consists in analyzing a
situation in terms of objective features and then finding a situation-
action rule which determines what to do.
I will argue that expert system builders fail to recognize the real
character of expert intuitive understanding. Expertise is acquired
in a five-step process: The BEGINNER does, indeed, pick out objective
features and follow strict rules like a computer. The ADVANCED BEGINNER,
however, responds to meaningful aspects of the situation which are
recognized as similar to prototypical cases, without similarity being
analyzed into objective features. At the next stage, the COMPETENT
performer learns to figure out a strategy and to pay attention only
to features and aspects which are relevant to his plan. The fourth
stage, PROFICIENCY, is achieved when the performer no longer has to
figure out his strategy but immediately sees the appropriate strategy.
Finally, the EXPERT, after many years of experience, is able to do what
works without facing a problem and without having to make any logical
calculations. Experts presumably do this by storing many whole situations
and associated actions in memory and responding to their current situation
in terms of its overall similarity to a situation already successfully
dealt with.
On the basis of this model one can see that expert systems based
on rules extracted from experts do not capture the expert's expertise
and so cannot be expected to perform at expert level.
A review of the successes and failures of various expert systems confirms
this analysis.