Author: Graham Laight
Date: 02:57:28 01/08/03
Go up one level in this thread
On January 08, 2003 at 02:58:07, Walter Faxon wrote:
>>>(1) If we want a venue to explore advanced AI in chess, let's change the rules
>>>for a new series of competitions: "limited search" computer chess (LSCC).
>>>Humans search about 2 positions per second (Anand claims he searches 5). Well,
>>>our computers today are pretty dumb so let's start with a maximum 100
>>>positions/second; this can be lowered later. The computer horsepower now used
>>>for searching would be applied to complex pattern recognition and other AI
>>>techniques. All program sources would be made public after each tournament,
>>>like the "RoboCup" robot soccer competitions today. This would serve two
>>>purposes: to prove nobody is searching faster than allowed, and to spread the
>>>wealth of knowledge. This would revitalize academic computer chess: everybody
>>>learns from everyone else and there is no point in competing unless you have a
>>>new idea or a much better implementation. Many more papers would be written.
>>>And any ideas developed could also be adopted by the fast searchers, as
>>>possible.
In the past, Ed has found that REDUCING the amount of knowledge improves Rebel's
Elo rating.
However - the fact that humans can play at SGM levels proves that a high level
can be achieved with knowledge alone. Presumably, you need a lot of knowledge to
get to the point where you're going to start getting rapid gains - and
presumably the knowldge v strength graph has some plateaus on it.
Knowledge (x axis) v Strength (y axis)
======================================
******
**
**
*
*
*
*
*
*
*
*
*
*
*
**
****
**
*
*
*
*
At the top of the graph, beyond a certain knowledge level, all games are drawn.
>>I would LOVE for this competition to happen. If enough people were willing to do
>>this (which, regretfully, is unlikely), then I'd happily take part!
>
>
>Someone has to go first; why not you? Write your program, making it as general
>as possible with sockets for various AI methods, do experiments, write a paper
>or two, then make it all public. Others can plug in their own ideas and
>methods, and boom! we have a small herd of pattern-oriented limited-search chess
>programs with authors eager to show them off.
>
>For one pattern-based approach to chess, you might revisit the 1979 PhD thesis
>of David E. Wilkins (See http://www.ai.sri.com/~wilkins/bib-chess.html). An
>impressive work, and the basis for a couple of papers in the AI Journal. A
>possible start would be to reimplement his logic in Prolog.
>
>
>>It might suggest it, but it falls far short of proving it. If you go to the SSDF
>>website ( http://w1.859.telia.com/~u85924109/ssdf/ ) and look at the games,
>>you'll see that games between stronger chess programs are more likely to end in
>>draws than games between weaker ones. This, together with the fact that it has
>>been proven that there's no forced win of material in the first 30 moves, proves
>>beyond reasonable doubt that chess is a draw with optimal play.
>
>
>I disagree. Stronger computers draw each other more often because each can
>avoid falling victim to the other's tactics. They are not exploiting subtle
>advantages; they don't know how. They stumble around a while and if neither
>side happens into the winning line that neither side sees, they draw. Why else
>is a computer plus a competent human player (not necessarily a grandmaster or
>even a master) so much stronger than a computer alone?
All other things being equal, the golfer with two different clubs to choose from
will beat the golfer who is only allowed one.
Anyway - your answer nae explains why weaker chess computers don't also get a
high proportion of draws. The games are all there to be seen at the SSDF site.
>Even if there is no forced win of material in the first 30 moves (determined by
>a full 60 ply search!?!?), that doesn't mean that one side doesn't have a
>winning advantage. There are books of openings analysis where (for example) a
>variation is refuted by a line where one side queens one move ahead of the
>opponent, on move 75.
>
>Do you have a URL for this "no forced win of material" claim?
The claim was made here at CCC by the author of SOS (I think his name is Huber).
He ran his program overnight with an ultra-fast "material only" evaluation.
>Because we don't need Go for a research platform. Every AI problem that is
>found in Go is also found in chess and we can force ourselves to confront these
>problems if we throw away the crutch of brute-force search!
In Go, you have a large board with a large number of pieces - so the mathematics
of pattern searching soon become unmanageble - one has to start guessing at
which patterns to search for.
In chess: suppose there are 32 pieces, and you're looking for patterns of groups
of 5 pieces. That's 32_C_5 combinations - or 32!/(5!*27!) = 201,376.
If one expands this to look for patterns in groups of 6 pieces, that's
32!/(6!*26!) = 906,192.
Suddenly, I see the possibilities of brute force reappearing! :)
Having said that, one would wish to search only for the existence of likely
patterns in the current position - possibly using techniques like
knowledge-nets. Also, combinations of patterns are likely to be needed.
-g
>-- Walter
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.