[computer-go] Re: Amsterdam 2007 paper

[Rémi Coulom]

Yamato wrote:
> Rémi,
>
> May I ask you some more questions?
>
> (1) You define Dj as Dj=Mij*ci+Bij. Is it not Aij but Bij?
>     What does this mean?
>   
Yes, it is ! Thanks for pointing that mistake out.
> (2) You have relatively few shape patterns. How large is each
>     pattern?  5x5, 7x7, or more?
>   
I use radius 3,4,5,6,7,8,9,10, according to the distance defined in 
Table 1. This looks very much like those used by de Groot and the 
Microsoft guys, with some very small differences. With a radius of 10 
according to my distance, the most distant point is 5 vertices away from 
the center.

I did not make big efforts to learn more patterns, and bigger ones, 
because I found that they do not improve the playing strength. It 
improves prediction rate a lot, but not playing strength. Crazy Stone is 
not significantly stronger with patterns of size 3 to 10 than it was 
with patterns of sizes 3 and 4 only.

That may be because it is not efficient to use knowledge in the widening 
algorithm that is not available to the random simulations. Also, large 
patterns are useful only in the opening, not in the middle game where 
most crucial tactics take place.
> (3) You say "the nth move is added when 40*1.4^(n-2) simulations
>     have been run." How did you determine these numbers?
>   
I tried plenty of alternatives and kept what produced the best strength 
against GNU Go. Remarkably, I found that the same formula produces good 
strength, whatever the size of the board. The alternatives I tried were 
linear widening (really does not work), and changing the values of 40 
and 1.4. Performance is not very sensitive to those values. I tuned them 
when I was using less clever patterns, so it may be that they are not 
very optimal.

Thank you very much for your feedback.

Rémi