From: Stephen Reed (reed@cyc.com)
Date: Mon May 20 2002 - 09:52:17 MDT
On Mon, 20 May 2002, Michael Roy Ames wrote:
> In reference to: "We are working hard to create a real AI."... any
> additional information and/or detail that you can provide would be very
> welcome here. The interplay of ideas, the discussion of problems and
> solutions, will help us all to improve our designs.
Based upon the "Good Old Fashioned AI" notion that a commonsense KB, and a
robust N-Order logical language is the foundation for an AI, Cycorp:
1. is investing in strengthening its deductive inference engine, adding
user-friendly knowledge acquistion tools in the Darpa RKF project,
2. is extending Cyc's understanding of structured knowledge sources in the
Darpa DAML project and in another government SBIR grant.
3. in the Darpa EELD project we are providing the reference ontology
(vocabulary) that will enable collaborator's text information extraction
tools to discover relationships among entities in the counter-terrorism
domain.
Our first product, Cyc-Secure, uses Cyc's planner to identify possible
ways an adversary can compromise a government or enterprise network. This
product is in use at a strategic military facility.
Although the shortcomings of semantic nets are well described in Elizer's
writings, we are reviewing the entire Cyc ontology, starting at the top as
it gets released into OpenCyc. We do make the distinction between
internal programming object and external referent. We acknowledge that
the average of 10 assertions per term is very far short of what an AI
needs to know about a concept.
What Cyc has:
1. Over 100,000 terms arranged in a single rooted generalization hierarchy
permitting fast subsumption reasoning.
2. Over 10,000 relationships arranged in generalization hierarchies,
permitting precise assertions of over 1 million facts in a hierarchy of
contexts.
3. A robust N-order deductive inference engine that has sacrificed the
ability to get an answer for certain, for the full expressiveness of
N-order predicate calculus vocabulary. That is, you can describe a set of
facts and rules to Cyc to answer questions, but Cyc may not be able to
answer your question.
4. Full support for functional notation. It is not necessary to name
every concept of interest. CycL's functional notation allows composition
to refer to concepts, without explicit reification (naming).
5. The GUI tools for Ontological Engineers (non-programming logic
experts) to manipulate and query the KB.
6. A growing set of GUI tools for Subject Matter Experts (SME's) to
interact with Cyc via clarification dialog.
7. A lexicon of 10,000 English Language terms linked to Cyc concepts,
allowing Cyc to parse some simple sentences, and to generate English
language sentences corresponding to each of its assertions.
8. An embedded version of WordNet 1.6, the well-known lexical database.
9. A robust blackboard for agent interaction.
10. A planner adapted from the open-source UMBC SHOP planner.
What Cyc does not (yet) have:
1. Reinforcement learning.
2. Goal directed behavior.
3. Self-initiated tasks,
4. Robust NL understanding.
5. Fuzzy logic inference, although CycL can represent Fuzzy lexical terms.
6. Self-modifying code.
7. many other aspects of AI....
Regarding Seed AI, Cycorp researchers (naturally) believe that the KB is
the key (so that some time in the future):
1. Extend Cyc's vocabulary to represent its own behavior.
2. Introduce Cyc behavior governed by KB content.
3. Elaborate that behavior to create the Seed AI.
All of the above will be included in OpenCyc for you to examine, review,
and extend.
OpenCyc can inter-operate with other AI systems via its java api, and via
CycML - the serialization of CycL in XML. OpenCyc is freely
distributable, and you could embed it within a AI software system of your
own design.
One motivation for publishing OpenCyc is to encourage one reference
vocabulary for representing commonsense knowledge. In this forum I am
especially interested in the vocabulary required to share knowledge among
the AI systems planned or under development.
Regards,
-Steve
--
===========================================================
Stephen L. Reed phone: 512.342.4036
Cycorp, Suite 100 fax: 512.342.4040
3721 Executive Center Drive email: reed@cyc.com
Austin, TX 78731 web: http://www.cyc.com
download OpenCyc at http://www.opencyc.org
===========================================================
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT