"Fat fritz" stronger than Stockfish ?

[MikeB]

Fat Fritz is a derivative of some sort, obviously it cannot be a clone of Alpha Zero since that code was not shared, but probably Lc0, created by Albert Silver for ChessBase. It is a program publicly available on the ChessBase engine cloud and supposedly its running on pretty big hardware. I think that's pretty cool.

https://en.chessbase.com/post/fat-fritz ... th-is-that

[Master Om]

[d]1b2r1k1/1p3ppp/1B3n2/1N1p4/P2P4/1Q2pPPN/1P2P1KP/2q5 b - -

Leelenstein also finds 24. ... Qc4 in under a second.

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[Analysis "Allie v0.4"]
[AnalysisTime "51"]
[FEN "1b2r1k1/1p3ppp/1B3n2/1N1p4/P2P4/1Q2pPPN/1P2P1KP/2q5 b - - 0 1"]
[SetUp "1"]

1... Qc4 2. Qxc4 dxc4 3. Ng5 Nd5 4. Ba5 f5 5. f4 h6 6. Nf3 Re6 7. Be1 Bc7 8.
Nxc7 {(0:00:51) 172kN}{[%eval -785,10,Allie v0.5][%meval 51s]} *

[M ANSARI]

I think he was trying to point out that even with a weak laptop GPU, the NN engine was able to find a move that the AB engine with strong hardware did not see.

But that point isn't valid if the move it "finds" is weaker, as the posts above suggest. We've seen this issue before.

I really haven't analyzed the move, but the point here is that "Fat Fritz is stronger than SF" was false and nothing in the article suggests that. It has more to do with giving the chess analyzer a different angle of thought as AB engines would very quickly prune out a move they doesn't make any tactical sense on an AB engine's search horizon. Chess wise this has profound impact on human chess as 100% of chess GM's today use computer analysis to test out moves or ideas in their preparation. So I think the point of the article was more to show that with an NN engine you have a tool that can give you more options and more dimensions to ideas or moves you want to test out with your engine.

[jp]

I really haven't analyzed the move, but ...

You don't need to analyze it. Just look at the comments in this thread.

So I think the point of the article was more to show that with an NN engine you have a tool that can give you more options and more dimensions to ideas or moves you want to test out with your engine.

Forget about the thread title for now. (And yours is a charitable view of what the point of the article was. The article makes it sound like one NN is special, when the truth is it doesn't give anything good that other NNs can't give.)

My point is that we've seen in this forum repeatedly that NN engine moves people get excited over either are found by AB engines too or are just weaker than the best moves.

[M ANSARI]

Forget about the thread title for now. (And yours is a charitable view of what the point of the article was. The article makes it sound like one NN is special, when the truth is it doesn't give anything good that other NNs can't give.)

My point is that we've seen in this forum repeatedly that NN engine moves people get excited over either are found by AB engines too or are just weaker than the best moves.

I totally disagree. I think with NN engines there are some moves that AB engines simply don't see or won't play. Just look at how much SF has gained in strength since Lc0 came out. Of course with infinite time an AB engine should see even the best NN engine move, but in chess you don't get infinite time. Thus, an NN engine seems to use "experience" from previous games (or AI if you will) to play a move that just seems right without having to go through all the deep tactical intricacies. I think of NN engines use shortcuts to tactical depth or maybe a better term is that they "translate" a position as being better without going through all the tactical details. This is very similar to how a human player plays chess as even the strongest chess playing human has incredibly limited tactical ability when compared to even a weak AB engine. I mean what does "positional chess" even mean ??? For me it simply means translating complicated tactical positions into a simple set of rules that are better understood by humans. Of course AB engines do have some sort of AI already built in as they have parameters that can modify its evaluation to navigate a position without tactics. If you look at SF, it has gained a huge boost in strength as it learned from its losses against Lc0. You could say the gaps between AB engines and NN will close as the holes in AB engines get patched up. Looking at lost games by SF against NN engines, it does seem that losses are due to several holes in evaluation that SF repeatedly falls for. My guess is those holes will eventually be patched and SF will get stronger. Who knows maybe eventually every single hole will be patched up and the CPU count on processors and CPU efficiency is only going to increase with time. But until that happens you will still get moves that an NN engine finds that are not even on the radar of an AB engine search.

[jp]

I totally disagree. I think with NN engines there are some moves that AB engines simply don't see or won't play.

Don't just disagree abstractly. Please try to give examples of some best moves that NN engines see that AB engines don't. There have been many potential examples looked at in this forum, and none of them has ended up not seen or to be the single best move. There are also examples of claimed differences in evaluation of NN and AB engines, which weren't differences or were probably correctly evaluated by the AB engine, not the NN engine.

[Nordlandia]

This is a position that Lc0 outplays SF.

[d]3r1rk1/pp6/1bn1R2p/6pb/5p2/1NP2N1P/PP3PPB/3R2K1 w - - 0 8

[jp]

This is a position that Lc0 outplays SF.
3r1rk1/pp6/1bn1R2p/6pb/5p2/1NP2N1P/PP3PPB/3R2K1 w - - 0 8

Swap the colors (so SF has Black) with the same short time control and see whether SF outplays Lc0.

That was actually one of the examples I was thinking about. (Did you continue reading the thread you started on that position? We can continue further discussion in that thread.)

It's likely a draw with best play, though you wouldn't want to be the one trying to draw it. How can people object then to SF's evalution of 0.00? It's true that seeing 0.00 may not help the human much, but that's a problem of what info the software gives us, not a problem of an engine's evaluation.

In your thread, it sounded like some people wanted the evaluation to be -2.5 or similar, but that would just be a wrong evaluation. (Lc0's evaluation also suggested a draw.)

[Ovyron]

It's likely a draw with best play

<Snip>

How can people object then to SF's evalution of 0.00?

Because an evaluation of 0.00 isn't useful. An engine that shows 0.00 for all moves that are "a draw with best play" is useless, and in practice it'd play really weak.

It's like, 1.g4!? is likely a draw with best play, if Stockfish showed 0.00 for it nobody would be happy. Centipawn scoring should reflect the difficulty that the defending side has to find this "best play", if it's very difficult the evaluation should be high (so if you're the attacking side, this position is appealing and better than other moves.) If the NN can see this, and show it, but Stockfish keeps showing a 0.00, M ANSARI's point stands.

[jp]

It's likely a draw with best play
<Snip>
How can people object then to SF's evalution of 0.00?

Because an evaluation of 0.00 isn't useful. An engine that shows 0.00 for all moves that are "a draw with best play" is useless, and in practice it'd play really weak.

But you want it to show you as correct an evaluation as possible. If it is drawn with best play it should tell you that. It should not make you think it's a forced win. In this example, it's really refuting the human's false belief that it's a forced win, which (if it's true) is not at all "useless". (Really, in the example, it's not showing exactly 0.00, but people were complaining that it wasn't showing a huge score, when if anything they should be complaining that it wasn't showing exactly 0.00.) It also shows you variations, so it can tell you how it thinks you should draw. By your reasoning, tablebases are "useless" too.

You might want it to show extra info too, but as I already wrote above, that's a matter of what the program or GUI displays, not related to the engine itself. What you're really complaining about is not an engine or its evaluations, but the display of just a single centipawn score.

And in practice SF does not play really weak chess.