Computer Chess Club Archives


Search

Terms

Messages

Subject: Symbolic: Status report 2004.11.05

Author: Steven Edwards

Date: 13:10:43 11/05/04


Symbolic: Status report 2004.11.05

It's been over two months since the last status report.  My apology once again
is that health issues have prevented doing much coding at all in September and
October; however, I have been able to to more thinking on the future development
strategy.

There have been a few positive items.  One of these is that Symbolic now has a
semi-permanent residence on the ICC and has played many hundreds of games
against a good variety of opponents.  This in turn has helped expose a number of
programming errors that did not show up in earlier testing.  I am thankful for
all of those humans and programs that have taken the time to battle against my
modest offering.  At this point in time, Symbolic's toolkit has managed a
roughly 2400 Elo in blitz games.  I doubt that the toolkit will get much
stronger as I think it's almost as strong as it needs to be to support the
upcoming Lisp based cognitive search.

A particular nasty bug in the opportunity (ponder) search was revealed via ICC
play with losses due to time mismanagement.  It turned out that the clock values
and move counts handed off to the separate opportunity search thread were
sometimes twice what they should have been; an extra update was being performed.
 One can easily imagine the difficulties produced by an error like this.

Another problem concerned the expectation scores returned on occasion by certain
fast searches; they were somewhat random and way out of normal range.  This
affected several post search routines including the PV/score display and the
resigner code: a fix is in progress for this bug.

The search status return class (CTMetaSS) is being extended to handle better
specification of the reason for search termination.  An example: book move vs.
tablebase move.  This will be propagated along the post search processing path.

There appears to be a problem with the toolkit sometimes starting a regular
search from scratch even when the opportunity search is pondering on the
correctly predicted move.  This is now under investigation.

Symbolic does not at the present have a static exchange evaluator in its
toolkit.  I'l be experimenting with a capture/promotion search instead of a per
square swap off technique.  Although a capture search will be slower than a
simpler swapper, the extra time per node available in a typical cognitive search
will allow this extravagance.

I hope to wrap up nearly all of the C++ toolkit programming by the end of 2004
and then return to a focus on the Lisp code running on Symbolic's ChessLisp
interpreter.  I expect that the full system's performance will drop by about
1000 Elo when it debuts on the ICC sometime in early 2005 as it will likely take
a lot of time to get the main functionality of the cognitive search in place.

One concept I've been toying with is to supplement the usual pedestrian Lisp
symbol pattern matching with techniques based on software holography.  The basic
idea is to transform a chess position into a two dimensional image and then
"illuminating" this and comparing  it with selections stored in a holographic
library of tactical and strategic motifs.  The resulting matches, if any, can be
used to guide the cognitive search that in turns deploys the toolkit for plan
verification.  The calculations will require a huge amount of floating point
operations and I hope that the Altivec vector coprocessor on the PowerPC with be
up to the task.

Can software holography work in a competitive chess program?  Here's a
tantalizing clue from regular holography: Make an optical hologram of a two
dimensional image of a typed word (e.g., "chess").  Now, take a page of text and
illuminate it with a laser of the same wavelength used to produce the hologram.
Finally, look through the hologram at the page.  Guess what!  Each appearance of
the word "chess" shows up as a bright light!  The closer the match, the brighter
the light.  It's almost like magic.  And some believe it's also a direct analogy
of the manner in which a human mind does pattern matching.  Now we're talking
about real AI, not that phony BS that's been spread around so much and for so
long.



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.