Panda - UCI chess engine
Author: Sebastiano Rebonato-Scott
Panda 1.1 what's new?
A second release of Panda, making some nice elo gains due to improvements in the search.
Some of the more notable changes include:
SPSA tuning using weather-factory
switch to gravity history formula
counter move and follow-up move heuristic
more sophisticated razoring conditions
improvements in late move reductions (conditions and reduction formula)
pawn structure correction history
switch to fail-soft search
time management based on proportion of nodes on pv move
However, these major changes probably only amount to about half of the elo gain, with the rest being from several smaller improvements.
Looking ahead, I think there's a lot of elo to be gained by improving Panda's neural network because:
Panda's current net was only trained on ~100M positions
These positions were scored by a much weaker version of Panda (pre version 1.0)
I don't think I'll have much time to write code in the near future, but if I can leave my laptop running in datagen mode for a while, I might be able to make another release.
Panda 1.1 - tests against other engines
Gauntlet vs 3200-3300 CCRL 40/15 (LTC)
I tested Panda against:
Nalwald 19 (3307)
Komodo 8 (3258)
Svart 6 (3242)
Princhess 0.19.0 (3224)
Lunar 0.2.1 (3200)
--------------------------------------------------
Rank Name Elo +/- nElo +/- Games Score Draw Ptnml(0-2)
1 Nalwald19 127.59 40.22 209.87 60.19 128 67.6% 39.1% [0, 3, 25, 24, 12]
2 komodo-8 63.12 47.20 82.82 60.19 128 59.0% 34.4% [2, 12, 22, 17, 11]
3 svart6 -2.71 38.89 -4.22 60.19 128 49.6% 45.3% [3, 15, 29, 14, 3]
4 Panda -14.12 20.07 -18.98 26.92 640 48.0% 38.4% [31, 76, 123, 68, 22]
5 lunar -49.18 43.29 -69.68 60.19 128 43.0% 40.6% [8, 17, 26, 11, 2]
6 princhess -63.12 46.54 -83.98 60.19 128 41.0% 32.8% [9, 21, 21, 10, 3]
--------------------------------------------------
Based on this result, I'd estimate Panda 1.1's strength at around 3240.
Jim Ablett compiles:
Comments
Post a Comment