[Pachi] self-play

Vlad Dumitrescu vladdu55 at gmail.com
Fri Aug 5 22:59:45 CEST 2011


On Fri, Aug 5, 2011 at 21:41, Matthew Woodcraft <matthew at woodcraft.me.uk> wrote:
> Vlad Dumitrescu wrote:
>> Is there some configuration that I should have been using, or
>> something? I just used the command-line from the README
>> ./pachi -t =1000 policy=ucb1,playout=light,prior=eqex=0,dynkomi=none
> That command line is an example of disabling many of Pachi's features,
> so it gives an unusually weak player;

I meant that - I presume that a small change will be easier to measure
for a weaker player.

> I expect that has something to do
> with it (sometimes they seem to be passing before all the territory is
> settled).

Maybe, but why does the estimate of the score depends on it being done
by one or the other of the players? The same program should estimate a
board the same regardless of who it plays as, shouldn't it? Even if
the situation is unsettled or the evaluation is wrong, the same error
should be made every time it sees the same board.

> If you add ',pass_all_alive' to the end of the command line, they'll
> play the game out until nearly all the dead stones are off the board;
> that should help them agree on the score.

Thanks, I'll try that.


More information about the Pachi mailing list