------------------------------------------------
v1.11.1
Changed : notation mode, was 3pts per victory. Now, takes into account the score difference at the end of the game. reverted back in 1.11.1
Changed : mutation probability skewed in favor of diminishing the size of the player (on average).
Bug fix : added some exception handling in the mutation process to prevent the program from crashing.
Bug fix : don't crash anymore on the first launch.
Changed : added komi rule of 6.5.
There have been many discussions on the computer go mailing list about whether to take into account the score at the end of a game, rather than simply the win/loss position. With MC players, there is strong evidence that taking into account the score makes the players weaker. This is because they may pursue more risky lines of play which lead to a big win, rather than safe play that wins by 0.5. I think the same should be true in a population of neural nets: if you reward big wins over small ones, this kind of player is more likely to emerge. In the end, you get the behaviour you reward. The goal is to win games, so that is what you should use in your point scoring system, in my view.
ReplyDeleteBy the way, that raises a couple of other questions:
are you playing the games with komi?
how come you have/had a points score for a draw? this should not be possible.
Keep up the good work. I had a lot of trouble with crashing, so will try this new version.
Oliver
Very fine, thank you for your advice.
ReplyDeleteHere you are with version 1.11.1 :
http://9giga.sfr.fr/n/50-17/share/LNK641149a8452353746/
(added komi = 6.5, moved back to the old scoring rule, no point for tie which can not happen anymore)
I mean, there it is : http://9giga.sfr.fr/n/50-17/share/LNK957249a847c08f251/
ReplyDelete