[computer-go] thoughts on 100,000 cgos games

[alain Baeckeroot]

Le Lundi 8 Mai 2006 05:12, David G Doshay a écrit :
> 
> As in my last post, I am not convinced that your statement must
> be true. It could also be that the board just evolved into a game
> that confuses the evaluation functions, and that we are likely
> to make our mistake very soon. Thus my comment that I agree
> that this can be used to point out a likely place to inspect ...
> but it still must be manually by a human.

A quick way to spare energy is to use selected games, only the _lost_ games
against _much weaker_ opponent (400 ELO below)
=> we are sure that it is our blunder that lose the game, and not opponent
strengh. (i find terrible move in gnugo like this, even one bamboo joint
being cut ! yes this is a bug ;)

Then `gnugo -replay ... --showscore` will very likely show a point when the
game was won and become lost, so we must dig before (maybe 3 moves is enough :)

This should help to eliminate the weakest part of the program.

I am curious to know which "parameters" could be tuned by automated learning
in SlugGo. As far as i rememeber, one must find some "parameter(s)" that have
significant impact on the game, and minimise some cost function wrt
the parameter(s), so need either stats on huge number of simulations or
a rather regular behavior of "params +--> cost(params)".

Alain