[om-list] OM version 0.02

[Thomas L. Packer at home]

Hello OM People

    I highly recommend "Artificial Intelligence: A Modern Approach" by
Russell and Norvig.  It is supposedly the best introductory AI text around,
and I believe it.  It adds unity to the field of AI by expressing all the
disparate ideas people think of as part of AI in terms of agents.

    As one example, I was amazed when I came to the part about decision
theory as applied to AI and designing intelligent agents.  This is my very
own pet theory, something I didn't think anyone else thought about, and I
then find that it is already well established in the field of AI, though
expressed
in slightly different words than I use.  In my words: actions should be
based on the correct combination of epistemology, ontology and axiology.  In
their words, The Principle of Maximum Expected Utility: "An agent is
rational if and only if it chooses the action that yields the highest
expected utility, averaged over all possible outcomes of the action."
Considering all the little parameters that can still be tweaked inside this
principle, I consider this to be a unifying principle of AI, in addition to
the idea of intelligent agents, and something that can apply to any
"intelligent" program written, even OM.  In fact, it should probably apply
to people's decisions as well.

    I wish I had gone to a school that actually had a good computer
science program earlier.  Even though BYU's CS department doesn't do a whole
lot with knowledge modelling and such things, it is still a good enough
school that I have learned a lot of useful things about what I want to do
for a career and/or research.  The issues seem a lot clearer when you
study from better sources.

    For example, (an example that is even more relevant to OM): I never
quite understood the issues involved in trying to model arbitrary knowledge
in a useful way.  Here is just one issue I have learned recently:

    If you want to invent a modelling language that is more expressive than
first order logic (predicate calculus), which is the standard from which few
people are diverging very far, then you will also need to make it usable by
also inventing a *complete* and *sound* inference procedure, something
analogous to what people use in such systems as Prolog and theorem provers,
i.e. something analogous to Modes Ponens or Resolution.

    Complete means that any idea that is true and can be expressed in your
knowledge base and which is consistent with your knowledge can be inferred
from your knowledge base.  --  And then you have to deal with Gödel's
Incompleteness Theorem which says that this is actually impossible if you
have a sufficiently expressive language (on that includes the principle of
"mathematical induction").

    If you want to try to beat this game by shifting from analytical
inference to synthetic inference (to induction from deduction) and want to
learn knew information, as in using machine learning, then you have to come
to terms with the "No Free Lunch" Theorems which say that it is impossible
to write a program that is sufficiently general-purpose so that it can learn
arbitrary information.  That is because the only way to induce generalities
from a finite amount of data is to make assumptions (machine learning calls
these "learning biases" and every learner has at least one).  These
assumptions can be true and useful for some areas of information, but they
will necessarily be false for others.

    Sorry to sound so pessimistic, but I have started to view my original
goal, and the goal of OM, as possibly being too idealistic -- and perhaps
impossible.  But I will continue to think about it.  And we should certainly
keep in mind the goal of usefulness: what do we want to do, and what will we
be able to do, with the knowledge once it is all in the box?

    My currently planned approach to the problem of modelling knowledge is
to tackle the input problem: natural language (a good source of knowledge):
make a program able to learn language and express the information it reads
as a knowledge base written in a language that can represent arbitrary
information.

    But now that I think about it, I have already decided that there is an
equivalence between inference and definition.  That is to say, there is a
correspondence between the two ways of inferring new information from old
(through induction and deduction) and the two ways of representing new
knowledge in terms of old (through intensional and extensional definitions).
This strongly suggests to me that it may be impossible to represent
arbitrary information, which definitely says something bad for OM if this is
right.  In fact, it may say that my goal at developing an interlingua is
doomed to failure.  I am not sure.

    I am totally re-thinking my research plans right now, trying to find a
more practical research goal -- one that is more likely to be possible, but
one that is still interesting and useful.

    We will see.
tomp

~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Omnia apud me Mathesis fiunt.
www.Ontolog.Com

It is a paradoxical but profoundly true
and important principle of life that the
most likely way to reach a goal is to be
aiming not at that goal itself but at
some more ambitious goal beyond it.
  -- Arnold Toynbee
~~~~~~~~~~~~~~~~~~~~~~~~~~~~


----- Original Message ----- 
From: "Luke Call" <lacall at onemodel.org>
To: <om-list at onemodel.org>
Sent: 2004.Feb.25/Wed 06:46
Subject: [om-list] OM version 0.02


I have released OM version 0.02 to the web site:
http://www.onemodel.org/devel/index.html

It still doesn't do much but adds minimal support for relationships and
some code cleanup. The relationships stuff is in the data model but only
partial UI support so far, and not really well tested. I thought I'd
post something before I add minimal "action" support (actor, action,
acted upon), so one could enter the equivalent of "we ate dinner" in the
system. Someday the action could represent a script which updates info
in the model (or even mini-simulations), but probably not right away.

I still need to go back & review some of Mark's earlier emails.

It's glacial, but progressing.

I hope all of you and yours are very well.

If anyone ever looks at the code & has feedback that would be great, but
no worries. :)

Best,
Luke

_______________________________________________
om-list mailing list
om-list at onemodel.org
http://six.pairlist.net/mailman/listinfo/om-list