perm filename FORUM[AM,DBL]1 blob
sn#415906 filedate 1979-02-02 generic text, type C, neo UTF8
COMMENT ā VALID 00003 PAGES
C REC PAGE DESCRIPTION
C00001 00001
C00002 00002 Professor Feigenbaum has given you a good feeling for what rule-guided
C00012 00003 2 PROPERTIES OF "DISCOVERY TASKS"
C00013 ENDMK
Cā;
Professor Feigenbaum has given you a good feeling for what rule-guided
expert reasoning systems are about: their design and the scope of their
applicability. He has really presented to you a whole "theory of
intelligence" in a nutshell <<SL: 3 parts>>; that theory says:
1. Human cognitive tasks can be cast as searches, as explorations wandering
through some problem space, usually toward some goal.
2. We are guided in these searches by a large collection of informal
rules of thumb, which we call heuristic rules, or just heuristics.
3. We access relevant heuristics in each situation, and follow their advice.
Intelligence is the ability to zero in on a solution effectively, despite
the apparent size of the search space.
That's it. It sounds plausible; in fact, it sounds trivial. And yet,
when we went off and built prgrams embodying this theory, they were
capable of expert performance at many tasks.
One of the big surprises was that this very same methodology is adequate
not only for scientific problem solving, but for scientific theory
formation -- for DISCOVERY tasks -- as well. And that's what I want to
talk about now.
There are two main properties of discovery that I want to communicate
today: <<SL: 2 aspects>> First, it is MECHANIZABLE. We've coded rules
which can guide a program as it explores some scientific domain, defining
new concepts, gathering empirical data, conjecturing relationships in that
data. Building and running such programs takes this investigation out the
realm of philosophy and psychology, and makes "the art of discovery" into
an empirical science.
Second, discovery is UBIQUITOUS. The same knowledge that guides you <<SL:
6 boxes>> to find a stairway in an unfamiliar building can guide you
effectively in defining a promising new mathematical function.
Let's look at this "ubiquity" in more detail, then we'll come back to
mechanizability at the end.
<<SL: Hier. of heurs>> Bear in mind the heuristics are not rules of
inference, but rules of thumb. They don't guarantee to preserve anything,
the way Modus Ponens preserves validity. They suggest plausible avenues
to consider, and prune away implausible ones. As we see here, the rules
appear at many levels of generality. Some of them are relevant in all 6
cases we considered before, some are quite specific to a particular field:
they are the repository of domain-dependent expertise as well as
domain-independent "common sense".
At first glance, each heuristic seems to have its own sphere of influence,
its own little domain of relevance, outside of which it's useless or
meaningless. A rule that talks about E. Coli can't have any bearing on
math research, can it?
If we look a little closer, we see that the heuristics can be organized
nicely into a generaliztion/specialization hierarchy. For instance,
"Study gene control signals across species of bacteria" is a special case
of "Study biol. mechnisms across species boundaries", which is a special
case of "Study the scope of a natural phenomenon". In that way, it does
make sense to talk about the first cousin of that E. Coli rule being some
math heuristic. Mathematicians have their stock of favorite
counterexamples, just as geneticists have their Drosophila and E. Coli,
and for the same reason: this general rule here: "In a new task, it helps
if your tools and subtasks are VERY familiar".
What I'm claiming is that the relevant heuristics, as a function of the
task they're being applied to, is a CONTINUOUS function. If you jump from
one task to another far away, it appears that the knowledge you need
changes completely. But if you deform the first task continuously into
the second, the heuristics will deform continuously also.
The molecular geneticist sits down to plan an experiment, say where he
wants to transfer a gene from one bacteria to another. The first thing he
does is to ignore the minor experimental steps, and build up a skeletal
plan. That wisdom is not unlike the knowledge you and I would bring to
bear when faced with planning not a genetics experiment but a long
automobile route: we would pull out a map and first ignore all the minor
roads.
<<<Everyday invention section of Ubiquity of Discovery>>>
<<<Scientific invention section of Ubiquity of Discovery>>>
<<<Judging sci. interest section of Ubiquity of Discovery>>>
This heuristic appears in texts on evaluating art and literature as well:
If 2 aspects of a work were perceived as distinct, but are suddenly
revealed to be related, then that's interesting. Charles Disckens relies
on this heavily, when characters turn out to coincide late in each novel;
Escher's art is dominated by this as well.
Enough about UBIQUITY; let's discuss the MECHANIZABILITY of discovery.
We built a program, AM, which had the definitions of a hundred finite set
theory concepts <<SL: concepts>> and a couple hundred heuristics, lik the
ones displayed up here. Among other discoveries, AM came across natural
numbers. Using the same line of resoning we did, AM defined prime numbers
and found out a few interesting things about them.
<<<Subset of the AM & Eurisko sections of Ubiquity of Discovery>>>
What we've learned from AM is guiding our current effort to write a program
which discovers new task-dependent heuristics as well as new concepts. It's
method is to represent each heuristic as a full-fledged concept, with slots
for its definition, origin, generalizations, and so forth. Other heuristics
can then operate on it, just as they would on any math concept.
We'll probably be working on this problem of "meta-heuristics" for a while
longer, but I'm still very excited about it: as this presentation must
have made obvious to you, I never met-a-heuristic I didn't like.
2 PROPERTIES OF "DISCOVERY TASKS"
MECHANIZABILITY
UBIQUITY