Raubfisch

[Stephen Ham]

Dear friends,

I look to you chess engine experts to educate me regarding chess engine strengths. That said, I'm surprised at the lack of posts regarding Raubfisch (RF).
https://chess.massimilianogoi.com/download/raubfisch/

Yes, the name is suspicious. It admits to being a derivative of Stockfish (SF) and Torpedo. It's also suspicious in that one of its two authors has the name Max Fehler (those who understand German will get the humor).

Nonetheless, some websites claim RF is slightly stronger than SF. This is very difficult to believe. So, I ran a 56-game match on my old and slow quad, using the same opening book, with a time control of: 40'/40 + 5'/15 + 2'+7"/game. Both engines played with default Contempt. The result was 7 victories for SF (Dec 16, 2018 iteration) versus 5 losses, which I don't see as statistically significant. This same result could have been achieved with two identical SF engines.

So, is RF identical to SF? SF has a size of 1.134MB while RF is 1.415MB. While both engines utilize time virtually the same, and analyze to the same average ply depth, their evaluations are different. SF's evaluation is volatile and often over-values positions in its favor. This is less true of RF; RF also displays minimal valuation volatility. Also middlegame move selection is slightly different. So, I conclude that the engines are very similar, yet different.

I analyzed all games post-match with SF (both Contempt factors turned off) and sometimes saw evaluations that differed significantly from the games. For example, I saw positions evaluated 1.31 centipawns in the game, but valued post-game by SF at 0.21 centipawns at the same ply depth. This 1.10 centipawn evaluation difference is far greater than can be explained by Contempt alone. The only other variable I can think of is RAM. For the match, the engines had slightly less RAM, while for post-game analysis they have slightly more. But RAM should not make that large a difference in the middlegame.

For the record, many games were decided in the endgame when both engines were living off of their 7"/move increment as they otherwise had no time left (many games exceeded 100 moves). As a result, some sloppy chess was played where it seems SF benefited during some time scrambles.

What are your thoughts, gents?

All the best,
-Steve-

[Guenther]

Dear friends,

I look to you chess engine experts to educate me regarding chess engine strengths. That said, I'm surprised at the lack of posts regarding Raubfisch (RF).
https://chess.massimilianogoi.com/download/raubfisch/

...


What are your thoughts, gents?

All the best,
-Steve-

http://talkchess.com/forum3/viewtopic.p ... 10#p691262

[corres]

Dear friends,

I look to you chess engine experts to educate me regarding chess engine strengths. That said, I'm surprised at the lack of posts regarding Raubfisch (RF).
https://chess.massimilianogoi.com/download/raubfisch/

Yes, the name is suspicious. It admits to being a derivative of Stockfish (SF) and Torpedo. It's also suspicious in that one of its two authors has the name Max Fehler (those who understand German will get the humor).

Nonetheless, some websites claim RF is slightly stronger than SF. This is very difficult to believe. So, I ran a 56-game match on my old and slow quad, using the same opening book, with a time control of: 40'/40 + 5'/15 + 2'+7"/game. Both engines played with default Contempt. The result was 7 victories for SF (Dec 16, 2018 iteration) versus 5 losses, which I don't see as statistically significant. This same result could have been achieved with two identical SF engines.

So, is RF identical to SF? SF has a size of 1.134MB while RF is 1.415MB. While both engines utilize time virtually the same, and analyze to the same average ply depth, their evaluations are different. SF's evaluation is volatile and often over-values positions in its favor. This is less true of RF; RF also displays minimal valuation volatility. Also middlegame move selection is slightly different. So, I conclude that the engines are very similar, yet different.
I analyzed all games post-match with SF (both Contempt factors turned off) and sometimes saw evaluations that differed significantly from the games. For example, I saw positions evaluated 1.31 centipawns in the game, but valued post-game by SF at 0.21 centipawns at the same ply depth. This 1.10 centipawn evaluation difference is far greater than can be explained by Contempt alone. The only other variable I can think of is RAM. For the match, the engines had slightly less RAM, while for post-game analysis they have slightly more. But RAM should not make that large a difference in the middlegame.
For the record, many games were decided in the endgame when both engines were living off of their 7"/move increment as they otherwise had no time left (many games exceeded 100 moves). As a result, some sloppy chess was played where it seems SF benefited during some time scrambles.

Raubfish is a derivative of Stockfish and this is admitted by its authors too.
Especially CC gamer often use specialized version of Stockfish.
Some of them are published and lot of them are not.
If you are satisfied with the play of Raubfish, use it.
The main issue of a derivative is it was not tested such thoroughly then the main version was.

[Guenther]

Raubfish is a derivative of Stockfish and this is admitted by its authors too.

...ridiculously admitted (meanwhile), because they illegally still don't release the source code,
which is of course easy, since the fake authors are fake people too...

[menniepals]

It's no secret where it came from.

[kramnik]

It's no secret where it came from.

where it came from?

[supersharp77]

It's no secret where it came from.

where it came from?

A very very strange "Bird" This Raubfisch...Tried them all from the earliest to the latest.For one thing the 'author' only produces one 'version' at the time.. (been doing it for a while yet never produced a w32 compile or a gen w64 compile)
I assume he is hex editing and modifying a Stockfish or Sugar engine..As far as "Torpedo" never heard of that engine and some claim it never existed at all..Sent a message to engine 'author' requesting additional compiles for some testing with no answer from the 'originator'...He does have a 'website', but I assume he is a very 'mysterious' fellow! Your guess is as good as mine..I have absolutely no idea! Got 40+ "Raubfisch' engines none of which work in Windows 10 x64 Intel nonpopcnt. Maybe one day I may get them to work down the road..But not today!

[Stephen Ham]

Dear gents,

I continue to test Raubfisch (RF) against Stockfish (SF) using the latest iterations, just out of curiosity.

Presently I'm running a 100-game match with a time-control of 30'+2" where both engines use the same opening book I created, leading to very sharp middlegames. After 85 games, RF leads with 8 victories to 7. Both play with default Contempt with Ponder off.

As stated in my initial post, their search factors (speed and depth) seem identical, although move selection sometimes differs. But the evaluation function seems quite different. Generally, I find RF's evaluation preferable - more accurate and stable.

So, whether RF is a clone or just a split off/fork version, it's a very strong engine that performs well. I'm still mystified why RF has more code (it's larger). Is it because it has yet to remove redundant SF code, or does this extra code improve evaluation? Anyway, it's all about results. RF's results generally match SF. Perhaps that's because it's so identical. A SIM test would be helpful here.

But as an ICCF GM with an important tournament starting in June, I'm considering using RF to help me in my correspondence chess games since I prefer its evaluation function over SFs. I also plan to start testing ShashChess as that seems an interesting engine.

All the best,
-Steve-

[Dann Corbit]

It's probably a Frankenstein made from Stockfish, Sugar, McCain, and a few other public projects.

I guess that if you use a bleeding edge Stockfish from here:

http://abrok.eu/stockfish/

Frankenfish will not do so well.

They're not quite so lazy as "Mr Dull Propeller" but I have seen historically that people who hide what they are and do illegal things are really not very smart. It does not take a lot of intelligence to cobble together some good ideas that are laying about and sewing them together to make something that looks new.

This string from the Raubfisch code:
"bbqnnrkr/pppppppp/8/8/8/8/PPPPPPPP/BBQNNRKR w KQkq - 0 1 moves g2g3 d7d5 d2d4 c8h3 c1g5 e8d6 g5e7 f7f6"

Is found in all the stockfish variants:
benchmark.cpp ( 86): "bbqnnrkr/pppppppp/8/8/8/8/PPPPPPPP/BBQNNRKR w KQkq - 0 1 moves g2g3 d7d5 d2d4 c8h3 c1g5 e8d6 g5e7 f7f6",

[Stephen Ham]

Hi Dann,

As stated in my post, I'm using the latest SF iteration, and yes, it's from that same "abrok" link you posted.

"Frankenfisch"...or whatever Raubfisch truly is, nonetheless leads Stockfish in my 100-game match after 88 games. While RF's 1-victory lead at this point is statistically immaterial, it is evidence that it's not clearly inferior to "bleeding edge" SF.

Like you, I'm highly suspicious of its origins, especially given the related joke names. Nonetheless, my primary interest is performance.

As an ICCF-player, I'm really looking for an analysis engine to help me with tactics in my CC games. So, that's why I created an opening book that leads to super-sharp middlegame positions. The hope is to determine which of these two engines has tactical superiority. I'll also retroactively examine some of my tactically sharpest CC games, where I've already completed extensive analysis, to examine the output of these engines.

In addition to a subsequent examination of ShashChess, I plan to also test asmFish and purchase the latest Komodo too.

All the best,
-Steve-