SUO: RE: Re: Architecture of an intelligent ontology development algorithm
Or, perhaps even more generally, the "better question" is how to combine
intentional, purposive design with natural selection operating on trial and
error? I think R&D in pharmaceuticals is a good working example of that
combination.
As to an upper ontology, I do not think that waiting for natural selection,
operating on a pool of candidates into which some process (analogous to
genetic mutation) introduces novel candidates or candidate-components, is
likely to succeed in any time frame meaningful to us. I think we need an
intentional, purposive design, tested for adequacy at each stage of its
evolution, against a wide range of low level ontologies taken from a diverse
set of concerns represented by working databases in various areas of
academic research and business function.
Top-down (intentional, theoretical), constantly tested and refined by
comparison with bottom-up (evolutionary, real world).
The testing and refining never stops, of course. So in theory, it could lead
to revisions in the highest levels of the ontology. Only through vacuousness
could the highest levels of our ontologies become immune to revisionist
pressures.
This, of course, is pure Quinean holism. But although, according to Quine,
even the laws of arithmetic are in principle subject to revision in the face
of recalcitrant experience, we count on them as being pretty stable. And
they have been. By the exact same token, I would expect a good upper level
ontology, once proven stable against a couple of dozen large, robust and
successful real world databases, to settle down into a stable state.
Nor does the benefit flow in one direction only -- lower level ontologies
helping with the development of higher level ones by being test cases for
their applicability. Upper level ontologies can also help us develop better
lower level ones, by revealing patterns in that lower level data that the
originating "subject matter experts" had never seen. I provided a brief
manufacturing example a week or two ago. Another set of examples come from
generalizing from a set of relational tables (or OO classes) to a common
supertype table (or class). Several vendor-provided "industry standard" data
models, such as IBM's banking model, define an INTERESTED-PARTY relational
table, subtypes of which include CUSTOMER, VENDOR, COMPETITOR,
REGULATORY-AGENCY. (Of course, this doesn't amount to very much, since
relational DBMSs support very little of the semantics of super/sub types. In
fact, in relational databases, they come to nothing more than one-to-one
relationships between the supertype and each subtype, optional for the
supertype, required for the subtype. So although the data model diagram with
its type hierarchy looks very sophisticated in a vendor's slide show, when
it gets down to implementation in a working database, it's much ado about
very little.)
Nonetheless, to summarize: it's top-down and bottom-up, design and
trial-and-error. If I have any content to add to this truism, it's this: the
process should be more top-down at the top, more bottom-up at the bottom,
but always both, at all levels. The influence works both ways. We
ontologists have something to add, something that reaches all the way down
into insurance claims processing databases, transportation freight bill
reconciliation databases, and grocery store shopping basket analysis
databases.
In particular, as I argued in an earlier email about manufacturing
databases, we should not think that the "real truth" is found in
functioning, real world, bottom-level databases. The ontologies they embody
are often confused, and the codebase and user knowledge of the system
substantially devoted to compensating for the confused ontologies. The whole
things are Rube Goldberg (Heath Robinson, for the Brits) contraptions,
usually and for the most part.
Tom
-----Original Message-----
From: owner-standard-upper-ontology@majordomo.ieee.org
[mailto:owner-standard-upper-ontology@majordomo.ieee.org]On Behalf Of
Jon Awbrey
Sent: Monday, August 25, 2003 5:45 PM
To: SUO
Subject: SUO: Re: Architecture of an intelligent ontology development
algorithm
o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o
[Reposting after 2 hours]
Rich,
I would have thought that a fairer summary of what's been said here
all summer long on many different threads is that there is really no
such thing as a hypothesis-free algorithm for discovery -- actually,
that's more like a summary of what's been discovered about discovery
over the last few thousand summers, but who's counting? So I think
that a better question might be something along the following lines:
How are concept-driven (analytic, axiomatic, rationalist, top-down) methods
and data-driven (synthetic, contingent, empiricist, bottom-up) procedures
best to be integrated in human inquiry, or in the reconstitutions thereof,
given that the distinction between analytic and synthetic is more
relational,
interpretive, or "situated" than it is absolute, invariant, or "essential"?
A start on answering that question might be to get a better analysis of the
similarities among and the differences between the various types of
reasoning
that need to be integrated. On that score, my advice would be: Read more
Peirce.
Jon Awbrey
o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o
Richard Cooper wrote:
>
> Since many of us seem to agree that a bottom-up
> algorithm could be used to produce the axiom
> set of an ontology through situated experience
> in the real world, I'm trying to draft some
> requirements for this algorithm.
>
> There is a very suggestive paper at
> http://jasss.soc.surrey.ac.uk/6/3/1.html
> "Discrete Agent Simulations of the Effect
> of Simple Social Structures on the Benefits
> of Resource Sharing".
>
> The paper desribes a simulation of agents in an
> environment somewhat like early human societies
> are thought to have evolved in.
>
> A similar approach could be used to measure the success of
> each strategy on the basis of how successful agents use that
> strategy. In a simulated environment, instead of a situated
> one, its easy to measure behaviors and organize them according
> to what works well and what doesn't.
>
> So in a situated environment, perhaps the algorithm can guess at
> axioms based on fragments of previous guesses that were successful.
> The so-called evolutionary algorithms could suggest requirements for
> monitoring the algorithm's behavior in the real world, measuring success
> and failure, and buliding a database of experience for process
improvement.
>
> So it seems to me that the process improvement concepts should be
> a top level ontology in an algorithm that learns still higher level
> axioms, while the WordNet concept set provides at least the words for
> communicating with real world people.
>
> Any thoughts on this subject?
>
> Thanks,
> Rich
o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o~~~~~~~~~o