Sergio Vieri second net is out

[Rebel]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

[chrisw]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

[Milos]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

The only thing that is rediculous is your knowledge about DNN training.
The loss happens because net is overfit. Problem with training of NNs as eval functions is that you can't just do validation as you do it with regular DNN training. A typical loss function is useless. You get smaller loss on your validation dataset but once you play games you get much larger loss. And there is nothing you can do about this. Coz eval is coupled with search so you can't just validate eval standalone.

[chrisw]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

The only thing that is rediculous is your knowledge about DNN training.
The loss happens because net is overfit. Problem with training of NNs as eval functions is that you can't just do validation as you do it with regular DNN training. A typical loss function is useless. You get smaller loss on your validation dataset but once you play games you get much larger loss. And there is nothing you can do about this. Coz eval is coupled with search so you can't just validate eval standalone.

Exactly what I meant, you show it. Typical “engineer” who knows nothing. You cannot explain in CHESS terms what it knows , doesnt know, or how or how not. Clueless nerd.

[Milos]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

The only thing that is rediculous is your knowledge about DNN training.
The loss happens because net is overfit. Problem with training of NNs as eval functions is that you can't just do validation as you do it with regular DNN training. A typical loss function is useless. You get smaller loss on your validation dataset but once you play games you get much larger loss. And there is nothing you can do about this. Coz eval is coupled with search so you can't just validate eval standalone.

Exactly what I meant, you show it. Typical “engineer” who knows nothing. You cannot explain in CHESS terms what it knows , doesnt know, or how or how not. Clueless nerd.

Lol chess terms. In the dictionary under demagogue they should have your name .
It's not shameful to be old and not understand new technology. But it's embarrassing trying to present it as some high level philosophy in the same time just because you don't understand it.

[dkappe]

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

Is the price you pay for freedom from bean counting.

Eavesdropping on discord;

1. These nets are trained 50% on score and 50% on game result (that can be adjusted)
2. The data generation inserts some random moves into the game
3. Apparently SV increased the maximum ply where this happens, causing the regression
4. Bean counters everywhere rejoiced

[chrisw]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

The only thing that is rediculous is your knowledge about DNN training.
The loss happens because net is overfit. Problem with training of NNs as eval functions is that you can't just do validation as you do it with regular DNN training. A typical loss function is useless. You get smaller loss on your validation dataset but once you play games you get much larger loss. And there is nothing you can do about this. Coz eval is coupled with search so you can't just validate eval standalone.

Exactly what I meant, you show it. Typical “engineer” who knows nothing. You cannot explain in CHESS terms what it knows , doesnt know, or how or how not. Clueless nerd.

Lol chess terms. In the dictionary under demagogue they should have your name .
It's not shameful to be old and not understand new technology. But it's embarrassing trying to present it as some high level philosophy in the same time just because you don't understand it.

Milos, neural net software is not exactly rocket science. Presenting as “overfitting” is no answer to anything, it’s a standard science-blinding which probably doesn’t even apply here, the Sergio nets are being constantly presented with new training data. Try getting your fundamentals straight before launching drama queen. Difficult and probably a bit too deep for for you, but lesson one in big boys world is not everybody is an idiot just like you were yesterday.

[Milos]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

The only thing that is rediculous is your knowledge about DNN training.
The loss happens because net is overfit. Problem with training of NNs as eval functions is that you can't just do validation as you do it with regular DNN training. A typical loss function is useless. You get smaller loss on your validation dataset but once you play games you get much larger loss. And there is nothing you can do about this. Coz eval is coupled with search so you can't just validate eval standalone.

Exactly what I meant, you show it. Typical “engineer” who knows nothing. You cannot explain in CHESS terms what it knows , doesnt know, or how or how not. Clueless nerd.

Lol chess terms. In the dictionary under demagogue they should have your name .
It's not shameful to be old and not understand new technology. But it's embarrassing trying to present it as some high level philosophy in the same time just because you don't understand it.

Milos, neural net software is not exactly rocket science. Presenting as “overfitting” is no answer to anything, it’s a standard science-blinding which probably doesn’t even apply here, the Sergio nets are being constantly presented with new training data. Try getting your fundamentals straight before launching drama queen. Difficult and probably a bit too deep for for you, but lesson one in big boys world is not everybody is an idiot just like you were yesterday.

Sorry but to have a meaningful discussion one needs to understand at least the basic concept how DNN training is done. You clearly don't understand it and are showing nothing but Dunning-Kruger here. But that's nothing new for you anyway.

[chrisw]

The last (2210) finally 59%, +63 elo above the official stockfish 11.

Code: Select all

SF-NNUE (popcount) vs Stockfish 11
henk-2706     54.7%   2020-07-19
sergio-1432   56.1%   2020-07-21 15:56
sergio-1907   58.9%   2020-07-21 19:14
sergio-2323   58.4%   2020-07-21 23:23
sergio-0359   58.8%   2020-07-22 03:59
sergio-0944   58.9%   2020-07-22 09:44
sergio-1153   58.3%   2020-07-22 11:53
sergio-1807   58.0%   2020-07-22 18:07
sergio-2210   59.0%   2020-07-22 22:10

I get a strong regression for 20200723-1134.bin

Same here, from 59.0% to 56.5%, a loss of almost 20 elo.

And nobody has the slightest idea why. The whole thing is kind of ridiculous.

The only thing that is rediculous is your knowledge about DNN training.
The loss happens because net is overfit. Problem with training of NNs as eval functions is that you can't just do validation as you do it with regular DNN training. A typical loss function is useless. You get smaller loss on your validation dataset but once you play games you get much larger loss. And there is nothing you can do about this. Coz eval is coupled with search so you can't just validate eval standalone.

Exactly what I meant, you show it. Typical “engineer” who knows nothing. You cannot explain in CHESS terms what it knows , doesnt know, or how or how not. Clueless nerd.

Lol chess terms. In the dictionary under demagogue they should have your name .
It's not shameful to be old and not understand new technology. But it's embarrassing trying to present it as some high level philosophy in the same time just because you don't understand it.

Milos, neural net software is not exactly rocket science. Presenting as “overfitting” is no answer to anything, it’s a standard science-blinding which probably doesn’t even apply here, the Sergio nets are being constantly presented with new training data. Try getting your fundamentals straight before launching drama queen. Difficult and probably a bit too deep for for you, but lesson one in big boys world is not everybody is an idiot just like you were yesterday.

Sorry but to have a meaningful discussion one needs to understand at least the basic concept how DNN training is done. You clearly don't understand it and are showing nothing but Dunning-Kruger here. But that's nothing new for you anyway.

Hmmm. When I need lessons from the little prince who never made it, I'll come and ask. Meanwhile, fault of the arrogant immature is to imagine that what you learnt yesterday is a new discovery that nobody else knows. Engineering ignoramus.

[dkappe]

If talkchess had a fragrance it would be called “Bile.”