Fat Fritz 2

[dkappe]

Dkappe's nets not clustering closely with SF's default doesn't invalidate the assumption, as their strength is also different.

What would one expect with what is essentially the same algorithm, just with different optimizations? Different results with the same eval would be a bug, no?

This comment betrays some serious ignorance about the nature of search functions in modern alpha-beta chess engines. Modern search functions are far cry from what most would consider an algorithm due to their reliance on a large number of ad hoc pruning, move ordering and reduction heuristics. These numerous heuristics are essential to achieving competitive performance and extremely relevant to an engine using an alpha-beta search's strength. While it is true that you will likely find many similar heuristics in top engines, top engines' search functions still vary significantly.

Different results with the same evaluation function are to be expected.

For a low fixed depth sim test, search function variety is less of a determining factor as history based move ordering heuristics are still "burning in" and some reduction/pruning heuristics are inactive at low depth. Some engines, such as Weiss (iirc), go so far as to perform a pure alpha-beta search up to the 6th search iteration.

I was half expecting this rebuke. So, does the pruning and such lead to a style that swamps the signal of an eval function? Or is it about the same as perturbing the eval by a few random centipawns? How would one test this?

[connor_mcmonigle]

Dkappe's nets not clustering closely with SF's default doesn't invalidate the assumption, as their strength is also different.

What would one expect with what is essentially the same algorithm, just with different optimizations? Different results with the same eval would be a bug, no?

This comment betrays some serious ignorance about the nature of search functions in modern alpha-beta chess engines. Modern search functions are far cry from what most would consider an algorithm due to their reliance on a large number of ad hoc pruning, move ordering and reduction heuristics. These numerous heuristics are essential to achieving competitive performance and extremely relevant to an engine using an alpha-beta search's strength. While it is true that you will likely find many similar heuristics in top engines, top engines' search functions still vary significantly.

Different results with the same evaluation function are to be expected.

For a low fixed depth sim test, search function variety is less of a determining factor as history based move ordering heuristics are still "burning in" and some reduction/pruning heuristics are inactive at low depth. Some engines, such as Weiss (iirc), go so far as to perform a pure alpha-beta search up to the 6th search iteration.

I was half expecting this rebuke. So, does the pruning and such lead to a style that swamps the signal of an eval function? Or is it about the same as perturbing the eval by a few random centipawns? How would one test this?

It's difficult to say. I would definitely expect Stockfish and Igel's similarity with the same network to diverge some with increasing depth provided the given position lacks a clear best move (this is a difficult property to insure). In any case, the evaluation function is still very important in similarity tests. However, as stronger entities are known to cluster closely in sim tests (Stockfish and Leela show unexpectedly high similarity for example), my prior expectation is that the FF2 net with SF's search will show high similarity with SF.

[dkappe]

It's difficult to say. I would definitely expect Stockfish and Igel's similarity with the same network to diverge some with increasing depth provided the given position lacks a clear best move (this is a difficult property to insure). In any case, the evaluation function is still very important in similarity tests. However, as stronger entities are known to cluster closely in sim tests (Stockfish and Leela show unexpectedly high similarity for example), my prior expectation is that the FF2 net with SF's search will show high similarity with SF.

Hmmm, deep search and high elo has a homogenizing effect? So this is an anti-style argument. I’d want to see some more substantial proof than shallow similarity tests.

[pohl4711]

A.Silver said, that the github net is weaker. And it is obvious, that this must be true. If not, nobody needed to buy FF2, so it would be crazy, if Chessbase would provide the strong commercial net on github for free.

hmm.. That would be just as crazy as selling a chess-enginge including network that is weaker than free SF versions ripping the same source, and claim it is stronger, but certainly not impossible... Can you help out with the MD5s of these network so we can actually verify that differnet rating lists are using the same net...

WIthout proper info on actual network versions used it all becomes a trust, mumbo-jumo issue.. Even for Pohls rating-list.

You can use the bench of FF2. As I wrote on my website:
Bench of Fat Fritz 2 commercial (Contempt +24) is 4025201, bench of Fat Fritz 2 GitHub is 4511552 (Contempt +24 is default)
So,if the bench is 4025201, it is the commercial net. If the bench is 4511552, it is the free github net.
The binary has no influence on the bench. Only the contempt has to be set to +24 for the first commercial binary. In the update1 binary and the github binary (both are the same) the Contempt is already 24 by default.

In my ratinglist, the FF2 entry is the commercial version (Contempt +24, commercial net). Bench is 4025201
The FF2 github-version will follow tomorrow: github-binary and github FF2-net. Will be a separate entry. Bench is 4511552

That is all perfectly clear. The three FF2-binaries (commercial (internal FF2 commercial net), commercial update1 (no internal net) and github) show the exact same behaviour/bench (if contempt is +24). And there are only 2 FF2-nets (commercial and github).

[pohl4711]

You can use the bench of FF2. As I wrote on my website:
Bench of Fat Fritz 2 commercial (Contempt +24) is 4025201, bench of Fat Fritz 2 GitHub is 4511552 (Contempt +24 is default)
So,if the bench is 4025201, it is the commercial net. If the bench is 4511552, it is the free github net.
The binary has no influence on the bench. Only the contempt has to be set to +24 for the first commercial binary. In the update1 binary and the github binary (both are the same) the Contempt is already 24 by default.

In my ratinglist, the FF2 entry is the commercial version (Contempt +24, commercial net). Bench is 4025201
The FF2 github-version will follow tomorrow: github-binary and github FF2-net. Will be a separate entry. Bench is 4511552

That is all perfectly clear. The three FF2-binaries (commercial (internal FF2 commercial net), commercial update1 (no internal net) and github) show the exact same behaviour/bench (if contempt is +24). And there are only 2 FF2-nets (commercial and github).

MD5 checksums:
commercial net: a89ad6de8979846ce3086b28ba0d606b
github net: 2ca7a3e70129e57a8af603a17a5d9905

[MikeB]

CEGT have published their updated lists now, and have both Stockfish and Fat Fritz 2 on their "Best Versions" list. So they are treating them as separate entities as well.

http://www.cegt.net/40_40%20Rating%20Li ... liste.html

Well them , you are both being misleading in your ratings. Two wrongs don't make a right. The code is is in excess 99.9% Stockfish, with the only different being the Net, Which is like having a different evaluate function.

My question is, is there an undisclosed conflict of interest. Did everyone at CCRL pay for their version of Stockfish-FF2 ? If you received a gratuitous copy, of FF2, then CCRL waas basically bribed to show it as separate. Be careful how you answer, the truth always has a way of coming out.

If anyone of you received a copy of FF2 without paying for you, CCRL is lacking independence.

[smatovic]

CEGT have published their updated lists now, and have both Stockfish and Fat Fritz 2 on their "Best Versions" list. So they are treating them as separate entities as well.

http://www.cegt.net/40_40%20Rating%20Li ... liste.html

Well them , you are both being misleading in your ratings. Two wrongs don't make a right. The code is is in excess 99.9% Stockfish, with the only different being the Net, Which is like having a different evaluate function.

My question is, is there an undisclosed conflict of interest. Did everyone at CCRL pay for their version of Stockfish-FF2 ? If you received a gratuitous copy, of FF2, then CCRL waas basically bribed to show it as separate. Be careful how you answer, the truth always has a way of coming out.

If anyone of you received a copy of FF2 without paying for you, CCRL is lacking independence.

C'mon Mike, that is a meany. I myself got sponsored long time ago with some Nvidia GPUs, yet my engine is OpenCL not CUDA.

--
Srdja

[Graham Banks]

If you received a gratuitous copy, of FF2, then CCRL waas basically bribed to show it as separate. Be careful how you answer, the truth always has a way of coming out.

If anyone of you received a copy of FF2 without paying for you, CCRL is lacking independence.

What a load of tosh.

[hgm]

My last message got removed (though I can't see what would have violated the charter).

[Moderation] Oops, it seems I messed up yesterday. There were some reported postings in this thread, and after those the discussion seemed to focus on who was a moron and who was a crook. So I moved the whole lot to the Moderation Archive.

But it appears that there was still a whole lot behind that, (on following pages), which I had failed to notice, and it was moved together with the offending postings.

Now I moved the whole lot back to this thread, only leaving out the postings that engaged in, or referred to the name calling.

Again, my sincerest apologies for this.

[smatovic]

ChessBase doesn't have to do anything anymore. Every 3 months they take the current Stockfish dev. Version, makes update to a Fat Fritz 2.x and claim Fat Fritz 2.x (and not Stockfish) got 15 Elo better! Great prospects at the expense of the Stockfish developers.

+1 bump*

*post got by error deleted by the mods.

--
Srdja