a response to anthony Cozzie opinion on rybka

[Uri Blass]

The most interesting thing in Strelka is the material imbalance lookup tables. These come from this:
http://home.comcast.net/~danheisman/Art ... alance.htm
http://home.comcast.net/~danheisman/Art ... _pawns.htm

It is perhaps not surprising that Larry Kaufman is on the Rybka team.

Although Mr. Kaufman published his results a long time ago, I do not think anyone picked up the idea to put this into a chess program until the Rybka author bothered to do it. It is clearly a very, very good idea.

I do not think that this idea is the main reason for the fact that strelka is strong.

Without material imbalance tables strelka is weaker but the main reason is simply the fact that it has wrong evaluation of pieces and overevaluate bishop and knight relative to rook and pawn.

I guess that if you drop the material imbalance tables and correct the value of pieces then strelka is not going to lose more than 30 elo rating points.

Uri

I do not know why Strelka is strong. I do know that the material imbalance idea is unique to Strelka and Rybka (so far). I guess that there are other unique ideas as well, but I did not diagnose them yet.

Material Imbalance is not a new idea. You can see commented out information in versions of Toga, and it exists in older versions of Glaurung, for example, using values from Kaufman's papers. Rybka/Strelka has refined that a lot, with a lot more terms and I assume data from games. How Vasik came up with the values is the mystery, but I certainly have played with the idea of analyzing game outcomes for score corrections based on win percentages. So maybe that is the basic idea.

I think it is worth about 100 ELO overall, but note the Strelka material tables have a lot of information in them already present in other programs (like KBK is a draw). With corrected piece values and Fruit-like rules for this stuff, my guess it is might be worth about 50 ELO.

My guess is that it does not worth for strelka so much because of the following reasons:

1)Strelka does not use them in optimal way
(I remember that there was some improvement in later versions thanks to improvement in material imbalance knowledge).

2)Strelka already performed well against glaurung2.0.1 without material imbalance(lost only by small margin) inspite of evaluating positions clearly wrong unlike fruit because strelka without the tables have clearly wrong value of pieces and I do not talk only about simple endgames.

I also remember that the programmer of Naum found that material imbalance tables gave him only 30 elo improvement so inspite of believing that they can help I believe that their value is overestimated and chess is mainly about search(one ply more gives a lot and in my tests strelka depth n+1 always won against fruit2.1 depth n when strelka depth n never won against fruit2.1 depth n(I only tested n<=10).

The difference is huge and the same poor strelka that lost at depth 10 29-11 against fruit2.1 won 22.5-17.5 when I only increased her depth to 11.

Uri

Edit:Another comment-Material imbalance is not a new idea but I doubt if it was tried by tables like rybka's tables before rybka.

Uri

[Tony]

Mr Hyatt,

you complained, that rajlich collected all the knowledge, you and others had to offer and then used it in his new commercial program adding some new techniques, he found himself. why don't you have a look on the rybka 1.0 code and check out, which incredible new techniques he found, that he "hided" from the chess programing community?

according to cozzie, there's nothing like that.

but I guess you will complain again, that he used your new ideas to built the latest rybka. I really wonder, why crafty is 400+ points weaker than rybka?

Because Robert spends a lot of time teaching others.

Some appreciate this, some don't and some take advantage of that.

For 90% of the programmers here, the first SMP code they ever saw was Crafty.

Tony

I really had appreciated if an expert like you would have given a better defense for Bob against some "Horst" who simply addressed impolitely. More, he obviously even confused Strelka with Rybka...

I did, but it got deleted by the moderators every time.

Tony

[Graham Banks]

I really had appreciated if an expert like you would have given a better defense for Bob against some "Horst" who simply addressed impolitely. More, he obviously even confused Strelka with Rybka...

I did, but it got deleted by the moderators every time.

Tony

Tony - referring to some members as firstly stupid people asking stupid questions and then as silly people asking silly questions was not acceptable. That's why those posts were removed.

Regards, Graham.

[Thomas Gaksch]

Hi Mark,
do you know or anybody else how strong Strelka is compared to other engines? Is Strelka in any rating list?
If i read the discussion i always think Strelka is extremely strong. Is that true?

Thomas

[Thomas Gaksch]

Thanks Gabor,
wow, that is quiet strong.
Do you mean with newest version the version which is distributed with source code?

Thomas

[Uri Blass]

Hi Mark,
do you know or anybody else how strong Strelka is compared to other engines? Is Strelka in any rating list?
If i read the discussion i always think Strelka is extremely strong. Is that true?

Thomas

The newest Strelka is about Toga 1.4 beta5c 4CPU strength.

My guess that the 32 bit exe is at similiar level to Toga 1.4beta5c 1 cpu.

Based on the comments in the source there is 60 elo improvement relative to 1.8

1.8 has 2870 ccrl rating
see http://computerchess.org.uk/ccrl/4040.l ... t_all.html

It means that 2.0 32 bits has 2930 ccrl rating that is already better than
Rybka 1.0 64-bit 2921 +16 −16 68.7% −127.2 35.5% 1415

If you add 40 elo that 2.0 can earn by 64 bits then it means 2970 ccrl rating.

Toga 1.4 beta5c 4CPU has 3010 ccrl rating in this list.

Based on the CEGT we have the following
31 Toga II 1.4 Beta5c 4CPU 2902 23 23 456 52.6% 2884 48.2%
40 Toga II 1.4 Beta5c 2CPU 2884 18 18 893 50.9% 2877 40.4%
46 Toga II 1.4 Beta5c 1CPU 2866 18 18 876 57.5% 2813 37.4%
48 Rybka 1.0 Beta x64 2865 14 14 1596 60.6% 2790 38.0%

Uri

[ernest]

Yes, and the 64-bit version.

Where is it possible to download a 64-bit version of Strelka?

[revengeska]

I agree with you mostly in regards to the academia side of writing engines. However, Mr. Cozzie has looked at the Strelka code and believes there's nothing new or revolutionary in the code, it's just super optimized for search. I really haven't looked at the search code for either, so I really can't tell you how Crafty compares to Rybka beta in this regard.

[bob]

I agree with you mostly in regards to the academia side of writing engines. However, Mr. Cozzie has looked at the Strelka code and believes there's nothing new or revolutionary in the code, it's just super optimized for search. I really haven't looked at the search code for either, so I really can't tell you how Crafty compares to Rybka beta in this regard.

I have already seen descriptions of a couple of ideas I am going to try. Nothing remarkable, but something that will shorten an already short execution time in part of the evaluation. But it would probably take far more than an afternoon to wade through any program and verify that there is "nothing new at all"... Some ideas are often not visible because of how they are implemented, until a line-by-line analysis is done...

[ernest]

Maybe this is it: http://www.zshare.net/download/636253098b62b1/ but I don't remember where I found it.

Thanks a lot, Gabor, this is it!
The Strelka_2_0.rar I downloaded only had the 32-bit exe.

I found that, on the start_position, the 64-bit Strelka is 50% faster than the 32-bit.
Not bad! (but on that position, Rybka 64-bit is 85% faster than the 32-bit)