On Tue, Mar 6, 2012 at 5:00 AM, Stefan Kaitschick < [email protected]> wrote:
> Bots love to throw in useless ko threat type moves now and then. > Sometimes its just a wasted threat, at other times it loses a useful > liberty too. > Why do even strong bots show this "stupid" behaviour? > > Lazarus had that problem too and before that my simple bot had it really badly because it did not have the benefit of a tree search. I don't know if discounting the move is a reasonable solution or not, perhaps it is. But I think that if you can solve it in the playouts you solve it in general. You almost need to determine in the playout how to refute a stupid ko threat but I don't think that is simple at all. Don > Ko threat type moves have one thing in common: a big payoff if the > opponent answers incorrectly. > If the first response tested is wrong, then the ko threat will start off > with a very good winrate. > The correct response will be found fairly soon, but by then the original > move will have a winrate bonus. > This bonus is small, but it is there. > The problem is, that the winrate degradation, that a useless ko threat > causes, can also be small. > So if the (erroneous) bonus outweighs the (real) loss, the bot will > conclude that throwing in the ko threat is good. > To remedy this, I think the winrates of early replies that prove > ineffective, have to be discounted from the parent node. > I know its not as easy as it sounds, because the winrate also directs the > search. > I'm not proposing a method here, just pointing to the problem. > > Stefan > > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
