My strongest NNUE net (yet)

[ChickenLogic]

I've trained a new 256x halfkp net which came within 20 elos reach of SFdev. Previous nets were at least 70 elo weaker than SFdev. TC was 1+1 and I used the 8 moves opening book from Fishtest.

You can get it here:
It is mostly trained on depth 8 fens. The earlier generations didn't even include draws. It really struggles with some endgames.

I plan on getting 1 billion or more depth 10 fens (including draws & at least 6 men tbs available to the engine). Any help getting there will be greatly appreciated.

[cdani]

Many thanks!
Exciting times
I'm testing it. I will post the results.

[lkaufman]

I've trained a new 256x halfkp net which came within 20 elos reach of SFdev. Previous nets were at least 70 elo weaker than SFdev. TC was 1+1 and I used the 8 moves opening book from Fishtest.

You can get it here:
It is mostly trained on depth 8 fens. The earlier generations didn't even include draws. It really struggles with some endgames.

I plan on getting 1 billion or more depth 10 fens (including draws & at least 6 men tbs available to the engine). Any help getting there will be greatly appreciated.

I put it and the half-k application in a folder, and tried to install in the Fritz-17 gui, but apparently there is something else I need to do as it won't load. I have a recent (1 year old) 8 core I-7 laptop, windows 10, with a good GPU (not needed for this I believe). What else do I have to do?

[cdani]

I put it and the half-k application in a folder, and tried to install in the Fritz-17 gui, but apparently there is something else I need to do as it won't load. I have a recent (1 year old) 8 core I-7 laptop, windows 10, with a good GPU (not needed for this I believe). What else do I have to do?

Put the net in "eval" folder under the folder of the engine, or change the EvalDir parameter.

[cdani]

Quick test 30 + 0.2 of your net (stnnuecl1) against the previous best net known for me. Yours is a bit better:

Code: Select all

   # PLAYER       : RATING  ERROR   POINTS  PLAYED    (%)
   1 stnnuecl1    : 2868.0   10.7    269.0     504   53.4%
   2 stnnue       : 2844.0   10.7    235.0     504   46.6%

Now testing it against the latest one of Additional Pylon halfkp 256.

[lkaufman]

I put it and the half-k application in a folder, and tried to install in the Fritz-17 gui, but apparently there is something else I need to do as it won't load. I have a recent (1 year old) 8 core I-7 laptop, windows 10, with a good GPU (not needed for this I believe). What else do I have to do?

Put the net in "eval" folder under the folder of the engine, or change the EvalDir parameter.

Thanks, that solved the problem. I'm getting just about half the nodes per second of Stockfish 11, is that typical? I also notice that the evals for unbalanced positions are even larger than the already large Stockfish 11 evals, but I suppose that's only a display issue that won't affect play. It does seem to show pretty sensible PVs; it's probably already stronger than SF 11 on an equal nodes basis if it's just 20 elo down at half speed.

[cdani]

I'm getting just about half the nodes per second of Stockfish 11, is that typical?

Yes. It's a neural net, very optimized, but slower than a typical eval function.

[FormazChar]

Highly recommended to adjust slowmover since the currently more instable eval negatively changes the tm code. In 5+0.05 using slowmover=100 vs slowmover42 losses more than 110 elo loss in one test.

[ChickenLogic]

I put it and the half-k application in a folder, and tried to install in the Fritz-17 gui, but apparently there is something else I need to do as it won't load. I have a recent (1 year old) 8 core I-7 laptop, windows 10, with a good GPU (not needed for this I believe). What else do I have to do?

Put the net in "eval" folder under the folder of the engine, or change the EvalDir parameter.

Thanks, that solved the problem. I'm getting just about half the nodes per second of Stockfish 11, is that typical? I also notice that the evals for unbalanced positions are even larger than the already large Stockfish 11 evals, but I suppose that's only a display issue that won't affect play. It does seem to show pretty sensible PVs; it's probably already stronger than SF 11 on an equal nodes basis if it's just 20 elo down at half speed.

Fixed node tests have shown over +70 elo (older nets, haven't tested this one). But it seems that the tm and the net aren't working well together. Also the search is basically unchanged so there probably are a lot of improvements that would be needed to get NNUE to the top. At least the fixed node tests would imply a higher elo in normal TCs but it doesn't happen as of yet.

[lkaufman]

I put it and the half-k application in a folder, and tried to install in the Fritz-17 gui, but apparently there is something else I need to do as it won't load. I have a recent (1 year old) 8 core I-7 laptop, windows 10, with a good GPU (not needed for this I believe). What else do I have to do?

Put the net in "eval" folder under the folder of the engine, or change the EvalDir parameter.

Thanks, that solved the problem. I'm getting just about half the nodes per second of Stockfish 11, is that typical? I also notice that the evals for unbalanced positions are even larger than the already large Stockfish 11 evals, but I suppose that's only a display issue that won't affect play. It does seem to show pretty sensible PVs; it's probably already stronger than SF 11 on an equal nodes basis if it's just 20 elo down at half speed.

Fixed node tests have shown over +70 elo (older nets, haven't tested this one). But it seems that the tm and the net aren't working well together. Also the search is basically unchanged so there probably are a lot of improvements that would be needed to get NNUE to the top. At least the fixed node tests would imply a higher elo in normal TCs but it doesn't happen as of yet.

I got an even score in 20 bullet games (1' + .5") vs SF 11 on seven threads each, which is consistent with your above result. This was with slowmover set to 60, just a guess as to what might be good. Perhaps testing at fixed movetime would show how they really stand now if we remove the time management issue. My sense is that this will become the best Stockfish in the near future.