Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: How much further to go in Man-Machine?

Author: Ed Schröder

Date: 14:50:24 07/16/00

Go up one level in this thread


On July 16, 2000 at 06:11:49, blass uri wrote:

>On July 16, 2000 at 03:38:21, Ed Schröder wrote:
>
>>On July 15, 2000 at 18:10:53, blass uri wrote:
>>
>>>On July 15, 2000 at 17:40:00, Ralf Elvsén wrote:
>>>
>>>>On July 15, 2000 at 16:58:40, Pete R. wrote:
>>>>
>>>>>To take a different tack, I'm not particularly interested in the debate about
>>>>>whether DJ is GM strength or not.  When I can set up a position on my home PC,
>>>>>and it can tell me as well or better than Kasparov can what the best move is and
>>>>>*why*, in terms I can understand, then there will be nothing important left to
>>>>>do in computer chess.  But that's still a long way off, past the point when a
>>>>>home PC can supply enough horsepower to have a program beat the World Champ in
>>>>>match play.  Playing good enough to win and understanding chess as well as a GM
>>>>>are two different things.  And frankly from a commercial standpoint I'm more
>>>>>interested in training software that can do the latter, rather than just beat me
>>>>>up.
>>>>>
>>>>>In terms of DJ's performance, the question I'm musing about is whether a top of
>>>>>the line 8-way processor general purpose computer may be sufficient to do the
>>>>>job of beating humanity, subject to some lucky or brilliant tweaks in evaluation
>>>>>code.  In other words, is the matter of coming up with better positional moves
>>>>>in blocked positions, thwarting wing onslaughts, etc. a matter of putting in so
>>>>>much *more* evaluation code that an 8-way server can't do the job?  DB got
>>>>>around this by having massive amounts of eval parameters, all done in special
>>>>>hardware.  But the recent performances, warts and all, of top multiprocessor
>>>>>programs begs the question of how much more horsepower is really needed.  This
>>>>>is simply speculation of course, like most of these topics, but only the
>>>>>programmers would have a feel for whether they need another 1000 eval terms, or
>>>>>just better tuning.
>>>>
>>>>I have looked at most of Crafty's evaluation. Let me first say that
>>>>this is no Crafty bashing. I like the program. That said, the chess
>>>>content encoded is very crude. It is heuristics which may or may not
>>>>apply to the position in question. None of the little pieces of evalutation
>>>>does a better job than an average clubplayer. It is in combination
>>>>with the search it becomes so powerful. Of course I realize it is very
>>>>hard to balance such amount of code and I would fail miserably myself
>>>>if I tried. I am *not* complaining.
>>>>
>>>>I think we need many more plies
>>>>before the program can "teach" us what a position is really about, or
>>>>alternatively a much more precise (and slower) evalutaion. Since Crafty is
>>>>playing in the same division as the best commercials, I don't expect them to
>>>>have much more sophisticated evaluations. If they have, it hasn't given
>>>>then any superiority.
>>>>
>>>>Let me take this opportunity to ask someone who knows: Has Ed Schröder
>>>>chosen to add "everything" in Rebel's evalution. I might have misunderstood
>>>>some of his statements. He seemed to be disappointed that more chess
>>>>knowledge could give a weaker program in comp-comp games.
>>>
>>>I think that the main reason is that bigger knowledge does not know important
>>>things.
>>>
>>>It may be better than the default in small things but I do not know about cases
>>>when there is a big difference in the evaluation of a pawn between knowledge=25
>>>and knowledge=500.
>>
>>Uri, you must be kidding... :)
>>
>>Rebel knows a lot more when [Chess Knowledge] is set to its maximum.
>>Here are a few examples where key-moves are found 1-3 plies sooner
>>pure on positional grounds. Position 1 & 4 are about king safety,
>>position 2 & 3 are instructive about the value of pins.
>>
>>r2q3r/pp1k1pb1/3p3p/3P3n/5Bbp/2N5/PPPQN1P1/2K2R1R w - - id POS (sac); bm Bxd6!;
>>r2q1rk1/pp2p1bp/2n1Ppp1/2pn4/3pNP2/6P1/PPPPQ2P/RNB2RK1 b - - id POS (att);  bm
>>d3!;
>>5r1k/ppp3pp/3bPqb1/8/3Pn1PP/4BN2/PP3PK1/R2Q1R2 b - - id POS (pin); bm Nc3;
>>r1bq1rk1/3nbppp/p2pp3/6PQ/1p1BP2P/2NB4/PPP2P2/2KR3R w - - id POS (ks); bm Bxg7;
>>
>>The list is endless. Used program Rebel Century 1.0 in case you want to check.
>>
>>Ed
>
>I will check this positions when I have time but here is a position when I was
>disappointed by chess knowledge=500(I used Rebel century1.2a)

In Century 1.2 the chess knowledge parameter by accident was cancelled.
Testing the chess knowledge parameter with Century 1.2 makes no sense.
That's why I used the 1.0 version.


>This is the relevant position from the game Junior6a-chess system tal:
>
>[D]r4rk1/pp1n1pbp/2p1p1p1/3nP3/2B2P1q/2N4P/PPPBQP2/2KR3R w - - 0 1
>
>Junior blundered by 14.Qg4 and lost a pawn.
>The reason is that it has not trapped knight code.
>
>computers like this move.
>They believe that white can get the pawn back by 14.Qg4 Qxf2 15.Bxd5 cxd5
>16.Nxd5 and do not understand that this line is not good because of 16...h5
>17.Qg1 Qxg1 18.Ne7+ Kh7 and the white knight is trapped(it is impossible to know
>by evaluation that the knight is trapped but it is possible to suspect that it
>is trapped and give a big positional score of a pawn for it).
>
>Junior6a saw it at move 15 and did not lose the knight but it lost a pawn.
>
>Chess knowledge=500 does not help Rebel to see that 14.Qg4 is wrong.
>The difference in plies here should be clearly more than 1-3 plies.

You are missing the point, perhaps I expressed myself wrongly. The point
is that [Chess Knowledge=500] knows more than [Chess Knowledge=100] and
the fact some tactical key moves are found on earlier plies is a good
proof. You can also check the increased score which I think is true for
the 4 examples I gave.

In the case of your Qg4 example I am not sure if [Chess Knowledge=500]
helps or not. Using the maximum setting isn't a magic stick that solves
everything as many many things still need to be programmed.

Ed


>The question is if there are cases when chess knowledge=500 help rebel to see
>clearly different evaluation and not only to be faster by 1-3 plies.
>
>Are there cases when chess knowledge=500 can see that white has one pawn
>advantage very fast and chess knowledge=25 cannot see it even if you search more
>1-3 plies?
>
>Uri



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.