patricia devlog

[Whiskers]

Exactly. Example: Each won (non-sac) game in the endgame (which old Patricia 3.1 could not win) lowers the percents of the won sac-games, because it is a win without a sac. This lowers EAS-score, of course, because all scoring is done with these percent-values.

I noticed this back when I tried testing against Komodo aggressive, both engines had under 50,000 EAS
Fun with EAS, if you pit 2 EAS engines against each other.

Code: Select all

                                 bad  avg.win 
Rank  EAS-Score  sacs   shorts  draws  moves  Engine/player 
-------------------------------------------------------------------
   1    110906  15.24%  13.33%  13.26%   72   Patricia-3.01  
   2     66328  03.49%  22.68%  26.43%   66   REBEL-EXTREME  

Both killing each other fighting for EAS points, Patricia wins.

As for the match score -

Code: Select all

No. Name           Win Draw Loss Unf.  Score Games       %
----------------------------------------------------------
  1 REBEL-EXTREME +1803 =1139 -105   *0 2372.5  3047   77.9%
  2 Patricia-3.01  +105 =1139 -1803  *0  674.5  3047   22.1%

Total Games:    3047
White Wins:     1117 (36.7%)
Black Wins:      791 (26.0%)
Draws:          1139 (37.4%)

World upside down.

[Whiskers]

Good progress - have reached just under 500k EAS with the same strength as Patricia 4. Patricia now has three networks and I’m going to test more because why not? It’ll be fun to have an infinity gauntlet of nets

update: reached 500k and am now working on strength

[Whiskers]

Strength has been a bit more annoying than I thought. I’m still only 30 ltc elo over Patricia 4 - the net is just dragging down the strength so much. Time for another round of experiments I guess…

[Werewolf]

Strength has been a bit more annoying than I thought. I’m still only 30 ltc elo over Patricia 4 - the net is just dragging down the strength so much. Time for another round of experiments I guess…

But Patricia 4 is already very strong, if aggression goes up and Elo stays the same as Patricia 4 - that's a huge achievement.

[Whiskers]

Strength has been a bit more annoying than I thought. I’m still only 30 ltc elo over Patricia 4 - the net is just dragging down the strength so much. Time for another round of experiments I guess…

But Patricia 4 is already very strong, if aggression goes up and Elo stays the same as Patricia 4 - that's a huge achievement.

Quite a few people's main experience with Patricia is watching her play at TCEC and CCC. I want her to be strong enough to score the occasional incredible win against other top engines. Sacrificing and winning against Stockfish would mean quite a bit more to me than, say, doing the same against an engine such as Willow.

[Whiskers]

I ran into a fun FRC datagen bug that took nearly a week to fix, but I can confirm that adding 50m DFRC data to 250m regular data leads to better results on small nets than training a net on 300m regular data in regular chess (not to mention FRC chess where it's over 100 elo stronger...)

[Whiskers]

I ran into a fun FRC datagen bug that took nearly a week to fix, but I can confirm that adding 50m DFRC data to 250m regular data leads to better results on small nets than training a net on 300m regular data in regular chess (not to mention FRC chess where it's over 100 elo stronger...)

One more idea that makes data stronger is to ensure that no opening is too unbalanced (my threshold is 400 centipawns or more) - if so, the opening is replayed until a suitable one is found.

I've seen it thrown around that using captures with a SEE score of less than 0 in datagen rather than filtering all captures out gained; however, this did not work in Patricia.

After a few more experiments I'm going to generate as much data as I can and get Gabe to train a few more nets for me, hopefully resulting in a stronger but no less aggressive engine. Then it's off to the search mines until Patricia 5...