perm filename LIGHT.2[2,JMC] blob
sn#049348 filedate 1973-06-14 generic text, type C, neo UTF8
COMMENT ā VALID 00002 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002 GENERAL LINE CONCERNING LIGHTHILL REPORT
C00007 ENDMK
Cā;
GENERAL LINE CONCERNING LIGHTHILL REPORT
1. Lighthill has misunderstood AI. Primarily he doesn't understand
what the basic scientific problems are. His ABC classification
shows that he regards the basic research part of AI as concerned
with the central nervous system - with neurophysiology and
psychology as he elaborates it.
2. In fact, the basic research part of AI is the study of
intellectual mechanisms abstracted from how they are realized in the
human or animal. This is partly for its own sake and partly because
intellectual mechanisms are more accessible by studying problems
than by studying physiology or the behavior of animals and people.
3. The experimental study of these mechanisms is important to
discover their nature.
4. However, AI is in bad theoretical shape, and there is no present
agreement on what the theory should be, although there is now pretty
good agreement that the main problem holding up progress is that of
how to represent the kind of partial information about the world
that everyone knows and how to represent the kind of information
about the world we get through our senses.
The bad theoretical situation is perhaps one of the reasons outsiders
misunderstand the field. It is especially easy for physical scientists
to misunderstand, because the things that might seem to work on the
basis of their experience do not in fact work. Witness the lack of
success of those who have tried to apply information theory, etc.
5. Some evidence that Lighthill has misunderstood:
a. He says that the quality of a chess program is determined
by the quality of its evaluation function. This was the guess that
Turing and Shannon and Bernstein made, but the disappointing
performance of Bernestein's program showed that it is mistaken. The
quality of a chess program is more determined by its choice of what
part of the move tree to investigate. It took experiment to
convince people that this is so. The parts of the game where the
evaluation function is primary such as the opening are the parts
present computer programs play best. Where they fail is when
successful play depends on having a chess idea that dictates what
lines are to be investigated in determining whether this idea leads
to an advantage. We can get much better chess programs when we can
formulate what a chess idea is and can first tell programs chess
ideas and then have programs generate chess ideas. Whether this
will happen in time to beat David Levy in 1978 now appears doubtful.
b. It took the robot projects to kill the statistical
pattern recognition and perceptron approaches to vision. These
approaches were concerned with classifying a pattern into one of a
finite number of categories. The approach doesn't work at all for
getting the visual information required for manipulation or driving.
Perhaps what is now known about description could have been
determined without manipulation-oriented experiments. In fact, it
wasn't.