On Wed, Sep 15, 2010 at 11:34 AM, Jeff Nowakowski <[email protected]> wrote:
> On 09/15/2010 07:09 AM, terry mcintyre wrote: > >> There are a fair number of joseki variations which have branches where you >> are >> not supposed to play X because you lose a capturing race. Initiate such a >> variation, leave it unfinished, move on to the next corner, repeat. >> > > The bot has to comply. It's one thing to speculate, it's quite another to > demonstrate over a series of games that the bot consistently gets into > capturing races that it then loses. I'd love to see this bot-killer strategy > clearly exposed -- not just occasionally, but with something like 80% or > greater reliability. I have an idea! First, some philosophy, please bear with me: Every player has weakness and strengths that define how good or bad he plays. Some of the games are going to be decided in your favor or against you based on this. If a human has a serious weakness he gets punished for it. That is no different for people than for bots. The idea that a human can adapt to this is somewhat myth, somewhat truth. If a player adapts to his weaknesses then it's just another way of saying he has improved in some way. Years ago someone wrote a book on how to beat Bobby Fischer at chess. It was an interesting book because it tried to analyze what you would have to do to maximize your chances - which openings to play, which style of game to aim for, etc. So I'm thinking that bot killer strategy is not special in this way. If you are an active tournament player and tend to face the same players you are interested in what it takes to beat them. Here is where the idea comes in: The issue with bots is that they are not trying to learn YOU. They are not readings books called "How to beat humans in Go." Therefore the authors must work this out for the computers. I think it's probably possible to trade one systematic weakness for another. I can do this in my chess programs if I want to. So then instead of having a fixed playing style a program could be made to adapt. But it would require that it is informed of who it's playing and maintain some kind of state for each player so that it can learn. So it is possible to trade one systematic weakness for another in a way that doesn't weaken either version too substantially in the overall sense? A related idea is that if you can simply provide a number of different styles to throw at people, it will be more difficult for them to adapt. I fear that you cannot substantially change programs in a way that makes them act fundamentally different. > > > _______________________________________________ > 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
