[computer-go] Crazystone patterns

[dhillismail at netscape.net]

> -----Original Message-----
> From: Rémi Coulom <remi.coulom at univ-lille3.fr>
> To: computer-go <computer-go at computer-go.org>
> Sent: Thu, 20 Sep 2007 10:22 am
> Subject: Re: [computer-go] Crazystone patterns
> Chris Fant a écrit : 
> > Does this mean that you need to calculate the Bradley-Terry 
> > probability for every legal move before selecting one on that 
> > probability? Isn't that expensive? Have you tried selecting only N 
> > legal candidates at random and then selecting one of those based on 
> > their Bradley-Terry probability distribution to save time? 

> I only keep light-weight local features in the random simulations. So, it is not expensive. A table of move urgencies can be > > > updated incrementally after each move: only a few change. In fact, my program became faster when I introduced 3x3 patterns > for the random simulations. I can pre-compute a lot of information indexed by pattern shape, that are useful to detect eyes > > > and move legality. From memory, Crazy Stone does about 5k playouts / second from the empty position on 19x19, on one 3 > > GHz CPU, which is "only" about 5 times slower than libego, if I remember correctly. 
>  
> Rémi 



> Chris Fant a écrit : 
> > Does this mean that you need to calculate the Bradley-Terry 
> > probability for every legal move before selecting one on that 
> > probability? Isn't that expensive? Have you tried selecting only N 
> > legal candidates at random and then selecting one of those based on 
> > their Bradley-Terry probability distribution to save time? 

> I only keep light-weight local features in the random simulations. So, it is not expensive. A table of move urgencies can be > > > updated incrementally after each move: only a few change. In fact, my program became faster when I introduced 3x3 patterns > for the random simulations. I can pre-compute a lot of information indexed by pattern shape, that are useful to detect eyes > > > and move legality. From memory, Crazy Stone does about 5k playouts / second from the empty position on 19x19, on one 3 > > GHz CPU, which is "only" about 5 times slower than libego, if I remember correctly. 
>  
> Rémi 




Suppose I can generate scores for all of the moves quickly enough. I still face the problem of quickly choosing a move biased by the scores. Tournament selection is fast, but that is a function of relative ranking of the scores, not the values of the scores. Roulette wheel selection gives me an answer, but it is slow slow slow, the way I implement it anyway. Can anybody describe a good way to do this?

- Dave Hillis





________________________________________________________________________
Check Out the new free AIM(R) Mail -- Unlimited storage and industry-leading spam and email virus protection.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20070920/74f46206/attachment.htm