[computer-go] EGC2007

[Rémi Coulom]

Jason House wrote:
> I believe the original paper used 32 simulations from each point as 
> part of its pattern...  But the spirit of it (as I understand it) 
> really is to bias the 1-ply move selection at the start of MC 
> searches.  (In Crazy Stone, just 3x3 patterns are used at deeper ply?)
>
> elife wrote:
>> I think CrazyPattern does not do any simulation at all. It just 
>> predict the move according pattern lib and Generalized Bradley-Terry 
>> model.

Crazy Patterns does run 63 simulations. But there is no tree search, and 
I do not believe that more simulations would make it play any better. As 
Cai wrote, Crazy Patterns simply computes a probability distribution 
with the generalized Bradley-Terry model. It runs a few simulations 
anyways, because Monte-Carlo owner is one feature of the Bradley-Terry 
model. But there is no tree search at all.

Rémi