JKAPLAN@SRI-KL.ARPA (09/26/84)
I recently learned that an article by John Dvorak criticizing our M.1
product in the San Francisco Chronicle 7/29/84 was reproduced and
distributed to the AIlist. This article presented a distorted and
factually incorrect picture of the Teknowledge product. The author
made no attempt to contact us for information prior to publishing the
article and as far as we know, has not seen the product. The article
appears to be based solely on information from a brochure, and
hearsay.
Based on the tone and content of the article, it was apparently
written primarily for entertainment value, and so we decided it would
not be fruitful to draft a formal reply. However, the AIlist might be
interested in a response. [I added a note to the original article
requesting such a response. -- KIL]
First about M.1 -
M.1 is a knowledge engineering tool that enables technical
professionals without prior AI experience to build rule-based
consultation systems. It is designed for rapid prototyping of
large-scale applications, as well as building small-scale systems. The
product includes a four-day hands-on course, extensive documentation,
sample systems, training materials, one year of "hot-line" support,
and maintenance.
M.1 contains a variety of advanced features. Some of interest to the
AIlist types include: certainty factors; a multi-window interactive
debugging environment; explanation facility; list processing; single-
and multi-valued attributes; variables; dynamic overlays of the
knowledge base during consultations; presupposition checking;
and automatic answer "completion". However, the system was carefully
designed so that it can be learned incrementally, i.e. the beginner
doesn't have to understand or use these features.
An initial CPU costs $12,500 (not $12,000 as stated in the article),
which includes training. Additional licenses costs $5,000 with
training, and $2500 without.
Strategically, M.1 fills a gap between mainframe- or lisp
machine-based tools for AI professionals, and a variety of less
sophisticated systems available to hobbyists.
Turning to the article -
Dvorak makes basically three points:
1. The program is overpriced for personal computer software.
2. The program gives bad advice about wine.
3. Expert systems are too complex to run on micros, at least with M.1.
Let me respond briefly to each point.
1. M.1 is not targeted to "personal computer owners" the way Wordstar
and VisiCalc are. M.1 is not intended, nor is it suitable for, mass
distribution. While M.1 can be used effectively without a graduate
degree in artificial intelligence, it is still quite a distance from
business productivity tools (such as Lotus 1-2-3) for non-technical
computer users.
Rather, it is a tool for technical professionals. We decided to host
the system on the IBM Personal Computer rather than the VAX or other
environments because (a) we believed this would be more convenient for
our target customers, and (b) it was technically possible without
compromising the product.
M.1 is priced consistent with similar systems that run on the IBM
Personal Computer, such as CAD/CAM tools, or modelling and simulation
packages. These systems typically appeal to a specialized audience,
and come with extensive training and support (as does M.1).
Our customers and the trade press understand the value of and
rationale for such systems. Some members of the popular and business
press do not. When we receive inquiries from these latter groups, we
explain the product positioning and provide appropriate references and
data points. We did not have this opportunity with Mr. Dvorak.
2. M.1 comes with a variety of sample knowledge systems, that
illustrate various M.1 features and suggest potential areas of
application. Skipping past extensive consultations in the M.1 brochure
with a Bank Services Advisor and a Structural Analysis Consultant, Mr.
Dvorak reprints an edited transcript of a sample system that provides
Wine Advice, in an attempt to ridicule the quality of the product.
In our brochure, the purpose of the brief wine advisor example is to
illustrate that the user's preferences can be taken into account in a
consultation, and that the user can change his or her mind part way
through a consultation. Initially, the user specifies a preference for
red wine, despite the fact that the meal contains poultry. The M.1
knowledge base naturally recommends a set of red wines. Mr. Dvorak's
version of the consultation stops at this point. In the balance of the
consultation, the user changes to moderately sweet white wines, and is
advised to try chardonnay, riesling, chenin blanc, or soave.
While it may occasionally provide controversial advice, the wine
advisor sample systems was reviewed by two California wine
experts before release, who felt that its advice was quite reasonable.
3. Regarding Mr. Dvorak's final point, he is simply wrong. Micros in
general, and M.1 in particular, are powerful enough to solve high
value knowledge engineering problems. Approximately 200 knowledge
base entries (facts and rules) can be loaded at any one time, and can
be overlayed dynamically if larger knowledge bases are required,
making the only practical limit the amount of disk storage. Through
the use of variables and other representational features, the language
is more concise and powerful than most of its predecessors. Practical
systems such as the Schlumberger Dipmeter Advisor and the PUFF system
at the Pacific Medical Center in San Francisco use knowledge bases
that could fit easily within the M.1 system without overlays.
For pedagogical purposes, we reimplemented a subset of SACON, a system
originally developed at Stanford University using EMYCIN, as a sample
system. SACON provides advice to structural engineers on the use a
complex structural analysis Fortran program. Our sample system
demonstrates that M.1 has sufficient functionality at reasonable speed
to accomplish this task. (The current version does NOT contain the
entire original knowledge base - time and project resource constraints
precluded our doing a complete translation. It includes all questions
and control rules, which account for about 50% of the original system,
but only about half of the judgmental rules, using no overlays. The
reimplementation can run the standard consultation examples from the
SACON literature.)
AIlist readers may be interested to know that M.1 has been selling
very well since its introduction in June. Our customers have been
extremely pleased with the system - many have prototyped serious
applications in a short period of time after taking the course, and at
a cost far below their available alternatives.
For more serious reviews of M.1, may I refer you to
Rosann Stach
Manager of Corporate Development and Public Relations
Teknowledge Inc
525 University Ave
Palo Alto, CA
415-327-6600
Jerry Kaplan
Chief Development Officer
Teknowledgerggoebel@water.UUCP (Randy Goebel) (09/30/84)
I've just read what amounts to an advertisement for Teknowledge's
M.1 software product. I can't believe there isn't something to
be criticized in a product that comes from such an infant technology?
I'd be interested to know what's wrong with M.1? Will Teknowledge
give it away to universities to teach students about expert systems?
Is SRI-KL using M.1 for anything (note origin of original message)?
On a lighter note, what is novel about a software system that supports
``variables?''
Randy Goebel
Logic Programming and Artificial Intelligence Group
Computer Science Department
University of Waterloo
Waterloo, Ontario, CANADA N2L 3G1
UUCP: {decvax,ihnp4,allegra}!watmath!water!rggoebel
CSNET: rggoebel%water@waterloo.csnet
ARPA: rggoebel%water%waterloo.csnet@csnet-relay.arpa