Updated Sunfish 2023

[thomasahle]

After the failure at the TCEC 4K event, I decided to run over sunfish and improve it bit.
The new version is less than 3KB when compressed using the new `build/pack.sh` tool, and about 150 ELO stronger than the previous version. (https://github.com/thomasahle/sunfish)

At 3min+2sec:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 sunfish 2023                  110      15    1602   65.3%   22.5%
   2 pygone 153                    -49      15    1600   42.9%   20.8%
   3 sunfish                       -58      15    1598   41.7%   24.1%

And at 30min+3sec:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 sunfish 2023                  155      41     241   71.0%   24.9%
   2 sunfish                         3      38     240   50.4%   24.2%
   3 pygone 153                   -160      42     239   28.5%   22.6%

(Here I tested with Pygone 153, https://github.com/scs-ben/pygone/ , a great minimalist python engine that I was destroyed by in TCEC. Partly because I forgot to use the fast pypy interpreter instead of the standard c-python.)

The main new features is a simple form of IID and a simple form of Futility pruning. Both done carefully to avoid search instability.
Still, the biggest improvement is probably a single line of code, which is discussed here: https://talkchess.com/forum3/viewtopic. ... f9#p940633

On the "interface side" it should now be easier to use sunfish with uci interfaces.

I was originally working on a NNUE version for the TCEC 4K, using the spare 1.5KB for a heavily compressed NN.
Unfortunately I didn't get it working in time. Maybe I'll have another go at it if another similar competition appears.

[thomasahle]

Now at 1000 games long time control:

Code: Select all

Rank Name                          Elo     +/-   Games   Score   Draws
   1 sunfish 2023                  176      20    1100   73.4%   22.8%
   2 sunfish                       -24      18    1102   46.5%   24.4%
   3 pygone 153                   -147      20    1098   30.1%   20.4%