From: Stephen Reed (email@example.com)
Date: Sat Feb 08 2003 - 20:33:29 MST
On Sat, 8 Feb 2003, Michael Roy Ames wrote:
> Dear SL4,
> Last month in the #sl4 chat channel, there was a discussion about
> competitive games and whether we should teach a seed AI, or Friendly
> Super Intelligence (FSI), about playing them.
Inspired by the recent Deep Junior vs. Gary Kasparov chess matches, I have
revisited (skimmed) the chess programming literature and critiques of
machine approaches to gameplay. Critics point to the game of Go
as too computationally complex for brute force look ahead. So knowing
very little about this game I read the simple rules and visited some Go
web sites for more information.
Here is a plan which I hope to get around to sometime:
1. Gnu-Go is open source Go software and interfaces with the Go game
communication protocols, so I can get Go realtime interface code there.
2. There are a couple of Java web client GUI's for Go that you can play on
the web and observe library Go games, and which interface to Gnu-Go or an application
I could program to comply with the Gnu-Go interface. This software gives
my potential application an interface onto the web Go gaming arenas.
3. A compelling critique of computer game play is that the best programs
have a large body of hand crafted heuristics which they employ. For
example in chess, the best programs have both a library of standard
openings and a hand crafted position evaluation function. I leave aside
whether the library of endgame positions is a heuristic or simply a form
of memoization. Critics say: "Why doesn't a smart computer just take the
rules and learn the rest from game play experience?". But programs taking
this approach have not turned out to be very good, and if I actually
pursue this task I'll have to examine these previous learning approaches
so as not to repeat their mistakes. I want to give a Cyc application
the rules of Go and a large body of commonsense gaming skill and see what
4. Although I am still just coding the initial stages of a NIST/RCS
hierarchical control structure as a Cyc application for automatic term
mapping, I have read and re-read James Albus' writings on this subject and
feel that his approach would lend itself to gameplay. I am especially
intrigued by the pattern recognition required for Go expertise.
5. The hierarchical control structure could easily be adapted for Go
gameplay I think. Each layer should have about ten times the planning
time horizon of the lower levels and I think that the smallest unit of
resolution would be a single move. So maybe three levels in the
hierarchy. Albus writes about sensory perception and the creation of
more general patterns from those sensed in the lower level. What would
be neat for me would be the chance to explore the pattern classification
algorithms available that could apply to the Go game-space. If Cyc is to
exhibit more intelligent behavior then it must learn to recognize salient
patterns, reifying them into symbolic concepts and extending Cyc's
ontology. On the planning side, plan operators would have to be formed
at first from a description of the rules and then expanded from learned
good groupings of simpler operations and their applicable gameboard
situations. Adversarial planning means recognizing the plans of your
opponent and this is something that we are just beginning to think
about at Cycorp.
6. The commonsense body of knowledge of gameplay could be applicable to
any adversarial situation and hence might get DARPA funding someday.
Especially if Cyc is able to learn how to play Go just from the rules, the
general gameplay knowledge and learning from experience.
7. I went ahead and ordered a Go board and stones - having them will be
a reminder to work on this.
Speaking of DARPA, Ron Brachman's visit to Cycorp in December went well
and we have hopeful expectations that his DARPA office may employ Cyc
as a standard software for an upcoming funded program. To further this
development, I would like to get some of my NIST/RCS style application
for automatic term mapping working in a couple of months for a compelling
demo of active Cyc intelligent behavior.
My java programming work in progress continues to be committed to
SourceForge but is out of sync with the released OpenCyc image. Cycorp
plans monthly releases of OpenCyc so I hope to get back in sync soon.
I gave a brief talk on the Singularity to Cycorp last month and it was a
hit. I plan a followup talk on the importance of Friendly AI and some
ways, as published by Eliezer, that it can be safely brought about.
Stephen L. Reed phone: 512.342.4036
Cycorp, Suite 100 fax: 512.342.4040
3721 Executive Center Drive email: firstname.lastname@example.org
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 - 04:00:41 MDT