Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Why Computer are not Grandmasters in strength.

Author: Chris Carson

Date: 09:26:14 06/11/01

Go up one level in this thread


On June 11, 2001 at 08:36:04, Mark Young wrote:

>On June 10, 2001 at 01:24:56, Robert Hyatt wrote:
>
>>On June 09, 2001 at 22:03:39, Mark Young wrote:
>>
>>>.
>>
>>
>>Sorry but that is a bit too sarcastic for my tastes.  I don't believe _anybody_
>>has said computers are not "GM-level" at fast time controls.  These two games
>>were 30 0.  We had a 30 0 event on chess.net two years ago with 4 computers
>>and 4 human GM players.  The 4 computers finished in the first 4 places.
>>
>>This is old news.
>>
>>40 moves in 2 hrs is something different.
>
>Why? Crafty may not be a 2500+ elo program, but there is a program called Deep
>Junior (You may have heard of it) that has played the best players in the world
>at 40 moves in 2 hrs at Dortmund and walk way with a rating of 2700+.
>
>Please explain why this result and others 2500+ results by other programs don't
>count.
>
>1. Junior was lucky?
>
>2. Did Mr. Ban payoff some of the GM's to lose to the Deep Junior?
>
>3. Other programs could not have achive such results so we can ignore this
>result?
>
>4. Computer programs still have weaknesses, and as we know human Grandmaster
>don't, as a result computers can not be considered Grandmasters in strength
>regardless on any results?
>
>5. Grandmasters have very fragile egos, and it is in some of our interest to
>placate the Grandmaster community to have continued access?
>
>6. Total results don't count as long as computers are still beatable and we can
>show losses to prove this point?

Mark,

Your points are right on target.  Many here (not me) will never say the programs
are GM strength unless they never have a below 2600 performance in only one game
and have 2600+ over thousands of games (very different standard than ELO
ratings).  That would mean that they would never draw a person lower than 2600
and would never loose to anyone with a rating lover than 3400 (2600+800).  We
are a long ways from that type of performance.  This is of course not how
ratings are established.  Ratings are an average (50% mark) for people or
computers not a 95% or 99% or 99.99% mark.  GM's have single game performances
that are very low (say a 2600 GM looses to a 2400 GM or 2400 IM) and the
performance for that one game is 1600, how could that be?).  GM's also have
performances in tournaments or matches every week that are +200 or -200 (several
games) or even +400 or -400 (fewer, but still there).  I would quote specific
human games here, but I do not want to offend anyone, so just take a look at any
issue of TWIC.

Programs can be beat, just as a GM can be beat (or any other person), the
programs today learn not to repeat mistakes so they loose much fewer games than
in the past.  Also, hw is much faster and this results in more consistent games
as well (unless there is a hw failure).  Besides HW and leanring, the programs
are just better than a few years ago.

Here is a look at 49 games played against 2500 or higher GM's, these are all the
games (except 12 DB games) that have been played against this level of
competition:

Program         HW      MHZ     Perf    Opp     W       D       L      Tot
DJ 6	        multi	8x700	2702	2702	2	5	2	9
Fritz 6/SSS	multi	4x500	2678	2545	6	4	2	12
P. ConNerS	multi	32x300	2663	2518	6	3	2	11
Rebel-Cen	P3	866	2598	2531	2	3	1	6
Rebel-Cen	K6	600	2537	2537	2	4	2	8
Socrates	multi			2545	1	0	0	1
Cilkchess	multi			2625	0	1	0	1
Ferrett	        multi			2630	0	0	1	1
Totals                                         19      20      10      49

That is almost a 2 to 1 margin of victory over very tough competition.  I do not
mean to pick on Ferrett.  Ferrett is one of the strongest programs and is a GM
strength.  Taking the one game against a GM and extrapolating that Ferrett must
me a 1830 (2630 - 800) player would be wrong in my opinion.

I see no great trend in these games that show that at 40/2 the programs can be
systematically beaten.  Fritz/DJ/Rebel are very well known and could should be
loosing not winning 2 to 1 if there was an easy way to beat these programs other
than "Ambush"

"Ambush" seems to be the best way to beat a program these days.  "Ambush" is
learning a programs weakness, then rest the program to default (and/or set
program to play at a faster time control) then replay the weakness.  I am
working on ways to reduce the possibility of "Ambush" in my program "Dallas".
It involves extensive learning and not allowing lerning to be reset.  A
discussion on "Ambush Defense" would be interesting for this group.  :)

Best Regards,
Chris Carson



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.