The next revolution in computer chess?

[corres]

As it seems the human programmer gets more and more pruned away.
Time to strike back.

Can we learn something from NN engines and improve AB engines with it ?

The two technique are so far from each other that there is nothing direct learning.
The only one opportunity to study the matches between Leela and Stockfish and modifying the stile of Stockfish such a manner that the style of Stockfish let be unpleasant for Leela.
An example: Stockfish is better in endgame, so it is beneficial for Stockfish with an early Queen-exchange to turn to endgame.
Or in closed game Leela is more successful than Stockfish, so the Stockfish`s handling of pawn should be modified so that the game should be opened or half-opened.
If you watch the play of Stockfish, you can see the effect of developers` modifications.

[dkappe]

I believe Naum is one of the better chess engines out in terms of positional evaluation. The only thing that holds it back is it's speed. Can you please do a tutorial in how to generate an NNUE net out of other engines eval? I've been following all the tutorials on Discord but none of them mentioned about using other engines as the teacher engine.

Training data can be converted to and from nnue binary training data format. The plain text format looks like this:

Code: Select all

fen 1rb2rk1/2p4p/2nq3p/p2p4/2pP4/Q4N1P/PP4PK/R4B1R w - - 2 20
move a3d6
score -164
ply 38
result 0
e
fen 1rb2rk1/2p4p/2nQ3p/p2p4/2pP4/5N1P/PP4PK/R4B1R b - - 0 20
move c7d6
score 180
ply 39
result 0
e
fen 1rb2rk1/7p/2np3p/p2p4/2pP4/5N1P/PP4PK/R4B1R w - - 0 21
move b2b3
score -185
ply 40
result 0
e
fen 1rb2rk1/7p/2np3p/p2p4/2pP4/1P3N1P/P5PK/R4B1R b - - 0 21
move c4b3
score 173
ply 41
result 0

That’s from some toga d8 data I’m generating. Score and result are from side to move’s perspective, so switch sign.

[smatovic]

I think since Peter Osterlund's experiment with Giraffe+Texel in 2017 the plain
potential of NNs in AB engines was shown...

http://talkchess.com/forum3/viewtopic.p ... 10#p719539

"Whether future hardware and software improvements will make it possible to run an ANN evaluator as quickly as a traditional evaluator remains to be seen."

Now we have the answer I guess, the right kind of tradeoff between NN structure
resp. size and AB-NPS on CPUs.

The LC0 project has its own value, but imo the MCTS-PUCT search was always a
kind of hack to make use of GPUs for ANN eval via batches.

Still plenty of space to explore, we just opened the can, or alike.

--
Srdja

[dkappe]

It occurred to me I should post a sample from a decisive game:

Code: Select all

fen r1bk1b1r/pp2p2p/nqp5/5p2/1PBP4/P1KQn3/R4NPP/1N5R b - - 2 17
move f5f4
score 518
ply 33
result -1
e
fen r1bk1b1r/pp2p2p/nqp5/8/1PBP1p2/P1KQn3/R4NPP/1N5R w - - 0 18
move c4a6
score -518
ply 34
result 1
e
fen r1bk1b1r/pp2p2p/Bqp5/8/1P1P1p2/P1KQn3/R4NPP/1N5R b - - 0 18
move c8f5
score 518
ply 35
result -1
e
fen r2k1b1r/pp2p2p/Bqp5/5b2/1P1P1p2/P1KQn3/R4NPP/1N5R w - - 1 19
move f2e4
score -506
ply 36
result 1
e
fen r2k1b1r/pp1bp2p/Bqp5/8/1P1PNp2/P1KQn3/R5PP/1N5R w - - 3 20
move a6c4
score -246
ply 38
result 1
e

There’s a random move in there since you want good positional coverage.

[Rebel]

...
NNUE engines don't need expensive GPU cards.

This is the only one real benefit.

Sure, everybody is wrong and you are right.

[corres]

...
NNUE engines don't need expensive GPU cards.

This is the only one real benefit.

Sure, everybody is wrong and you are right.

The future will decide about who is wrong and who is right.
In every cases I keep my NVIDIA cards.

[smatovic]

...
NNUE engines don't need expensive GPU cards.

This is the only one real benefit.

Sure, everybody is wrong and you are right.

The future will decide about who is wrong and who is right.
In every cases I keep my NVIDIA cards.

Hehe, dude, the future is now

http://ccrl.chessdom.com/ccrl/404/

Not sure if a such a thing happened ever before, #1 top entry on its first run @CCRL...

better invest in some AVX based CPU too

--
Srdja

[chrisw]

...
NNUE engines don't need expensive GPU cards.

This is the only one real benefit.

Sure, everybody is wrong and you are right.

The future will decide about who is wrong and who is right.
In every cases I keep my NVIDIA cards.

Hehe, dude, the future is now

http://ccrl.chessdom.com/ccrl/404/

Not sure if a such a thing happened ever before, #1 top entry on its first run @CCRL...

better invest in some AVX based CPU too

--
Srdja

Well, I am going down the contrarian path. NNs have their own exploitable problems too. For the time being.

[Dr.Wael Deeb]

...
NNUE engines don't need expensive GPU cards.

This is the only one real benefit.

Sure, everybody is wrong and you are right.

The future will decide about who is wrong and who is right.
In every cases I keep my NVIDIA cards.

Hehe, dude, the future is now

http://ccrl.chessdom.com/ccrl/404/

Not sure if a such a thing happened ever before, #1 top entry on its first run @CCRL...

better invest in some AVX based CPU too

--
Srdja

Wow ....

Thanks for the link ....
Dr.D

[smatovic]

...
NNUE engines don't need expensive GPU cards.

This is the only one real benefit.

Sure, everybody is wrong and you are right.

The future will decide about who is wrong and who is right.
In every cases I keep my NVIDIA cards.

Hehe, dude, the future is now

http://ccrl.chessdom.com/ccrl/404/

Not sure if a such a thing happened ever before, #1 top entry on its first run @CCRL...

better invest in some AVX based CPU too

--
Srdja

Well, I am going down the contrarian path. NNs have their own exploitable problems too. For the time being.

I mentioned it before, I guess this will be some kind of co-evolution between
expert-systems, handcrafted eval, and NNs. We see this in other domains too.

We will see which approach will perform better in the long run. But the paradigm
shift is already here, it came stepwise, NN eval vs. expert-system, if we take
a look closer, then the first steps were already the automated tuning methods of
our evaluation functions, NNs are just the next step in this line.

Further, if we consider that Reinforcement Learning offers us new paths, then I
would even go that far to call this beyond super-human-level chess, I would call
this RL ANN based eval "Trans-Human-Chess", cos it comes up with lines which
humans do not come up with, neither the machines they use to program by hand,
or alike, imho.

--
Srdja