[Pachi] Dynamic komi bug ?

lemonsqueeze lemonsqueeze at free.fr
Wed Jan 6 21:48:07 CET 2016


Hi All,

I'm trying to understand what happened to a game Pachi played on kgs 
recently:
     http://files.gokgs.com/games/2015/12/31/Yamoz-pachipachi-8.sgf

There's a big difference in strength and no handicap, so Pachi manages 
to kill 2 groups on a big scale and is winning by about 100 points. 
Clearly there should be no way to turn the game around at this point, 
but near the end around move 217 Pachi makes a massive blunder, gets one 
of his surrounding groups captured and loses the game (!)

It's playing with maximize_score to make it more human-like, and looks 
like it's trying to increase its margin even more instead of playing a 
safe move: If I load the position in gogui it does the right thing, but 
in the game winrate is 0.51 with extra komi 95 at this point, and if 
passed extra_komi=95 it does indeed go insane again =)

Not sure what's going on here, and how to fix this...
Does this look like a dynamic komi bug ?
Or it doesn't see the danger maybe ?

I've tried some initial debugging, board_estimated_moves_left() is 43 at 
this stage, so even with no_komi_at_game_end it'd still be a while 
before dynkomi logic gets turned off ...

Happy new year !
Cheers


More information about the Pachi mailing list