Re: [gnugo-devel] endgame move generator beginnings

[Arend Bayer]

Paul wrote:

> - endgame move generator beginnings implemented in endgame.c
> - endgame() in ascii.c renamed ascii_endgame()

First: I like the patch.

However, I tried to find a better way to test endgame patches than
running regressions on it (where not that many endgame positions occur).
I hacked twogtp.py (thanks for the Python version, Inge!) to allow
"endgame contests". By this I mean the following:

Load an almost finished game, and let the endgame be played out. Then
restart at the same position, revert colors, and replay it. Then take
the difference of the two results.
I hope this will give more unbiased results for endgame patches than
running regressions.

I tried this with 59 games. The version with your patch won 6 games, the
CVS version 5 games. (I wanted to play more games but messed up.) It
might be worth looking at the games that the CVS version won.

I've attached the modified twogtp.py. The details of the endgame match
are at http://www.uni-bonn.de/~uzsxtn/endgame_match.tgz.

(I produced this by using the game collection by Andries Brouwer
at http://www.cwi.nl/~aeb/go/games/index.html, starting the script
with --endgame 20, which means the game gets restarted 20 moves before
the end in the .sgf-record.)

Arend

twogtp-endgame.py
Description: Text document