For Stefan

In reply to : http://talkchess.com/forum3/viewtopic.php?p=936584#p936584

Used your advised tool, results:

REBEL 15.1

Code:

                                 bad
Rank  EAS-Score  sacs   shorts  draws    Engine/player
-------------------------------------------------------------
   1    180729  07.52%  55.31%  30.63%  "Komodo_14"
   2    169342  09.92%  48.35%  26.44%  "Stockfish_9"
   3    166412  20.27%  47.29%  17.25%  "Rebel_15.1"
   4    118873  08.26%  39.45%  24.91%  "Koivisto_6.16"
   5    116590  03.83%  45.90%  31.50%  "Berserk_6"
   6     93063  07.20%  45.08%  28.80%  "Seer_2.5.0"

REBEL-E380 (latest net)

Code:

                                 bad
Rank  EAS-Score  sacs   shorts  draws    Engine/player
-------------------------------------------------------------
   1    165598  23.58%  45.52%  18.63%  "Rebel-E380"
   2    156472  04.81%  57.22%  28.18%  "Komodo_14"
   3    153368  09.79%  55.67%  26.79%  "Stockfish_9"
   4     89778  03.49%  45.93%  32.65%  "Berserk_6"
   5     75533  09.83%  41.04%  28.17%  "Koivisto_6.16"
   6     66648  07.07%  40.91%  28.87%  "Seer_2.5.0"

Hi Ed, I agree with you that the better trained a NNUE is, the more balanced it will play, i.e. its aggressiveness will decrease.

I don't know if it is possible to train a NNUE to play more aggressively, maybe this would be possible if you "feed" it with gambit games or unbalanced openings.

Personally, I like programs that play more positionally and not speculatively. However, this is only my personal opinion.

Accepted

The EAS tool measures more than sacrifices, because aggressive play means not only playing sacs. But playing more short wins and avoiding bad draws (= draws before endgame and draws in games with (temporarely) material advantage), too.
And so we see, that the Rebel dev version plays some more sacs, but a little less percent of short wins and a little bit more bad draws. So, summarizing all these (that is, what my EAS tool does), the Rebel dev version is no progress in aggressiveness compared to latest released Rebel 15.1, instead the EAS score of both Rebels is nearly identical.

okay, thanks for clarification.

Admin wrote:

okay, thanks for clarification.

In the statistic output-file, below the ratinglist is the 2nd ratinglist with more detailed stats...
Sacs-percents from 1 pawnunit up to Queensacs and short wins from 40 up to 60 moves in percents.
This could be helpful for development of Rebel, I presume.