1992
	Response to Kyburg's Believing on The Basis of The Evidence
	a COMPUTATIONAL INTELLIGENCE discussion

	KYBURG AND VOLKSWAGENS





My understanding of probability is Kyburg's.  

I believe the reference class problem is the most interesting problem in
probabilistic reasoning (machine learning take note!).  I believe that
probability should mediate the fit of scientific theory to observational data
when predictive power and error must be traded against each other (plan
recognition beware!).  I believe the Bayesians are corrupt, or bankrupt, or in
any case, up to no good; I believe that we accept statements as firmly as I
believe anything.

My problem with Kyburg has to do with computational practice.

What are the implications of his philosophy to people who are building
programs?  What difference will it make when uncertain and inductive reasoning
is just a component?  Inexpressive language, resource-limited computation, and
poor modeling of preference are surely larger concerns for the knowledge
engineer.  Is Kyburg suggesting something that will pay new dividends in
artificial intelligence, or just a better metaphor?  I think Kyburg's essay is
merely an arcane description of current practice; it is an apology for how we
actually do things in our epistemological lives, unbeknownst to ourselves.  It
may be too painful for Bayesians and others to acknowledge, but it is what we
already do, at least when we are rational.

Elsewhere, Harman has been more successful at arguing why acceptance might
make computational sense.  A Bayesian knowledge engineer will condition on 
statements that are contingent, hence, not knowable with certainty, hence
accepted.  This is because we choose the input to the programs and we
usually choose a high level of abstraction -- we choose not to model
uncertainty at the level of robotic sensor inputs (and most Bayesian inference
engines are connected to diagnosis programs, not to robots).  Perhaps robots
and other image interpretation programs can be pure Bayesians and avoid
acceptance altogether.  The computational effort is immense when there is
unwillingness to build abstractions.  This explains why Cheeseman can remain a
pure Bayesian:  he is happy to crunch huge data sets.  The rest of the
Bayesians, for example Pearl, admit that they accept contingent sentences and
will continue to do so whether or not Kyburg notices.

Identifying reference classes is something that knowledge engineers don't much
worry about since they do not often work directly with sample data.  But
consider training connectionist networks on data sets, then using forward
propagation to predict some property for new inputs.  Shallow applications of
connectionism of this kind abound:  predicting bad weather, component of
speech, control of motion, handwritten characters.  A criterion of similarity
must be determined, whether by the magic of back-propagation or by the light
of reasoned methods.  The criterion must trade similarity against the desire
to bring as much of the past to bear as possible.  Consider, too, case-based
reasoning in a legal or problem-solving domain.  Ashley and Rissland's
"Waiting on weighting:  a symbolic approach to least commitment" perfectly
describes the non-Bayesian alternative Kyburg advocates:  prediction need not
be the result of weighting all past experience; some past experience might be
excluded and ``participate'' only with a weight of zero.  This is the same
question raised by reference classes.  Kyburg's own methods are worth
studying, implementing, and applying (I have done all three), but they do not
seem to lead to major advances over what is already being done.

Finally, Kyburg details how theory, measurement, and error interact.  This
is the jewel in Kyburg's philosophical crown; it is the legacy of a life lived
with probabilistic honesty.  Only Kyburg can give a good explanation of how
one comes to adopt some universal generalizations and not others, i.e., how
one theory or set of axioms is chosen over another.  Langley, Kelly-Glymour,
Kautz-Pollack, McCarty, Mitchell, Quinlan, and many, many more would do well
to understand why Kyburg has so much leverage when fitting rules to
observations.  Kyburg does not use familiar examples, such as curve-fitting,
when he could.  But even if they were to understand, they probably would not
change their practice.  Each has provided for the right tradeoff of error
against predictive power in his or her own way, or else has chosen to avoid
the issue.  To be enlightened by Kyburgian methods would only be to become
aware of an intriguing method, and then to dismiss it as too expensive
computationally.

So what are we to make of Kyburg's exotic machinery?  Ferdinand Porsche said
that if he were an automobile, he would want to be a Porsche.  There will be
programs that perform inductive tasks, and some programs will want to be
Kyburgian.  But they will always be a specialty breed, expensive to
manufacture and inappropriate for the masses.  Porsche's doom is Kyburg's as
well:  most of the time, a simple Volkswagen is good enough.