Texel & Velvet ... I like that so much!

[Frank Quisinsky]

Hi there,

Texel and Velvet play the openings mega aggressive with the final result that the move average from wins (without resign mode) are fantastic. And this in NN times vs. TOP-6-46 as opponents. Excatly this I like to see. Stats comes from FCP Tourney-2024, after round 5 of 50 (200 games).

TOP-TIPP for analyzing complicated tactcal opening positions!

Maybe the programmers can explain the secrets?

Code: Select all

33. Texel 1.10 NN                  (23+,   127=,   50-)  43.2%

          CATEGORIES   MOVE_AVE    GAMES    TOTAL  GAME_RATE
           all games         96      200
           white won         70       16
           black won         71        7
          white draw        101       73
          black draw        100       54
          white loss         97       11
          black loss         94       39
            all wins         70       23      200      11.5%
           all draws        101      127      200      63.5%
           all loses         95       50      200      25.0%

Code: Select all

36. Velvet 5.3.0 NN                (20+,   151=,   29-)  47.8%

          CATEGORIES   MOVE_AVE    GAMES    TOTAL  GAME_RATE
           all games         88      200
           white won         72       12
           black won         65        8
          white draw         81       84
          black draw         97       67
          white loss        105        4
          black loss        102       25
            all wins         69       20      200      10.0%
           all draws         88      151      200      75.5%
           all loses        102       29      200      14.5%

Best
Frank

[Frank Quisinsky]

Example more:

Velvet 5.3.0 NN around: 3300 Elo
Texel 1.10 NN around 3275 Elo
Stockfish 11 (without NN) around 3325 Elo (best short games stats in my older tourneys).

We can compare Stockfish with Velvet and Texel if I am using Stockfish 11 without NN.

Code: Select all

11. Stockfish 11                   (431+,   954=,   115-)  60.5%

          CATEGORIES   MOVE_AVE    GAMES    TOTAL  GAME_RATE
           all games         88     1500
           white won         79      271
           black won         82      160
          white draw         94      451
          black draw         91      503
          white loss         88       28
          black loss         89       87
            all wins         80      431     1500      28.7%
           all draws         92      954     1500      63.6%
           all loses         89      115     1500       7.7%

Can looking in my older tourneys with a lot of engines. Such statistics to "move-average wins" from Texel and Velvet with around a level from 3300 +-25 Elo I never saw before ... in the last 20 years I am looking on move average wins. Only Spark by Allard Siemelink had around 75 moves. Also Stockfish is not in the near. See the example from one of the Stockfish versions around with the same level Texel and Velvet have.

Code: Select all

13. Velvet 4.1.0 NN                (148+,   821=,   531-)  37.2%

          CATEGORIES   MOVE_AVE    GAMES    TOTAL  GAME_RATE
           all games         83     1500
           white won         71      106
           black won         79       42
          white draw         80      452
          black draw         83      369
          white loss         90      192
          black loss         86      339
            all wins         73      148     1500       9.9%
           all draws         81      821     1500      54.7%
           all loses         88      531     1500      35.4%

Can looking in older Velvet versions too ...
Allways the same ... white won 71 moves!!
Best others at this time are Stockfish, Komodo and Wasp.

A big different from Texel and Velvet to all others I tested. Different engines have strong results, like Komodo or Wasp, often Uralochka but not such strong results as Texel or Velvet.

Best
Frank

Velvet = 65 moves for black wins, simply unbelievable. I am looking in the next rounds of my tourneys, maybe random ... but I know all the stats from older Velvet versions. Thinking on Xiphos black move-average stats. Xiphos is playing with black much more aggressive as with white ... very rarely ... too bad that xiphos not longer in development.

[petero2]

Hi Frank,

Texel and Velvet play the openings mega aggressive with the final result that the move average from wins (without resign mode) are fantastic. And this in NN times vs. TOP-6-46 as opponents. Excatly this I like to see.
...
Maybe the programmers can explain the secrets?

I have also noticed that texel now seems to go for king attacks more often than in earlier versions that did not use NN evaluation.

I don't actually know why texel became more aggressive when I added NNUE evaluation. I know it is not because of the search code though, because texel 1.10 and 1.08 have exactly the same search code. All changes are in the NN evaluation.

I don't think it is because of the NN architecture, which is the same as in the original Shogi NN implementation, except that it also takes left/right symmetry into account.

I don't think it is because of the NN trainer, which does not use any special tricks. It is fairly standard C++ code using libtorch.

It might be because of the training data I use to train the NN. I generate training positions by using texel self-play games, but I don't use the positions that occurred in the games. Instead I search all positions in the games and randomly sample positions from the texel search trees. Those positions are then run through q-search to get more quiet positions, and those positions are scored using texel itself. Those quiet positions and the corresponding search scores are used in the NN training.

I cannot explain why this would create an aggressive engine though. This is unfortunately a common problem for neural networks. It is often very difficult to explain why they behave the way they do.

[Frank Quisinsky]

Hi Peter,

thanks for your answere, allways interesting!

With stats I can also not explain it.

At the moment I make on an other system different experiments with 40 moves in 8 minutes.
6 of TOP-10 engines have to play vs. 16 around 200 Elo weaker engines only. The six TOP-10 engines do not play against each other, means have to play games vs. the 16 around 200 Elo weaker engines only. Texel and Velvet have to play the same quantity of games vs. the same group of 16 engines.

Now I collected the first 75% of fastest win games for looking on "win-average".
At the moment the different is smaller to the others but again ... Texel and Velvet are on first two positions. Move average for fast wins (without resign factor) is around the same with 70 moves. Best TOP-5 engine have an average from 77 moves. If I am looking on 1. e4 opening systems only the different from Texel and Velvet to the others is 9 moves.

You wrote:
It might be because of the training data I use to train the NN. I generate training positions by using texel self-play games, but I don't use the positions that occurred in the games. Instead I search all positions in the games and randomly sample positions from the texel search trees. Those positions are then run through q-search to get more quiet positions, and those positions are scored using texel itself. Those quiet positions and the corresponding search scores are used in the NN training.

The way you go must be good!

After all, the secret is not to lose many fast games with all the aggressiveness.
I have the feeling that these engines also present me most interesting opening analyses (my main topic), generally most interesting analyzes with many pieces on board.

Before NN stats are much more simple.



Best
Frank

[Frank Quisinsky]

Forgotten ...
with 1. d4, 1. Sf3, 1. c4 opening systems Igel is in the near with a win-average from 73 moves!

Best
Frank

[DmitriyFrosty]

In my opinion, Texel play like a strong human player, has more humanistic style of play