Page 1 of 1
Strange pachi resignation
Posted: Thu May 29, 2014 1:53 am
by oca
Hi, I just did a quick game against pachi.
As I'm still a beginer, I started with 9 stones... the game was a bit too short in my opinion,there are still options for white is it
I started pachi with "-t =2000" maybe I must give him a bit more time

...
Re: Strange pachi resignation
Posted: Thu May 29, 2014 2:00 am
by moyoaji
What surprises me about this is that the resignation here feels human. If I were playing a game and made the silly mistake of letting my stone get captured like that I would ask for the game to be restarted (basically, a resignation). I think many human players would - that mistake is obvious, clearly the white player is tired/not thinking.
A computer shouldn't react like that - and I don't think it is feeling fatigued in any way.
I assume that the reason for the resignation is that the program isn't dealing well with the handicap stones. It is trying to analyze how it will win the game as if it was playing against itself. That is impossible for white to do. Perhaps you should lower the number of stones? Maybe start with 4 or 5 and see how the program does?
Re: Strange pachi resignation
Posted: Thu May 29, 2014 4:42 am
by Charlie
Pachi runs MCTS. I'd imagine Pachi resigns when its predicted probability of winning falls below a certain threshold. With nine handicap stones on the board, that probability must start out really low. Losing a hopeless push into your corner was clearly sufficient to push it below the resignation threshold.
Re: Strange pachi resignation
Posted: Thu May 29, 2014 5:25 am
by Uberdude
But to play handicap games most monte carlo bots use some kind of dynamic komi I think. So if we say 9 stones is worth 100 points and an average game lasts 250 moves then on move n give white 100*(250-n)/250 komi. But it's probably more complicated than that.