Re: [gnugo-devel] Mathematical Go: Chilling Gets the Last Point

[bump]

> I just browsed through this book by Elwyn Berlekamp and David Wolfe (I 
> am sure all of you know about it) and was wondering:
> 
> - does this have practical applications for GNU Go or for a Go AI in 
> general?
> 
> - if it does, how would the program recognize that it is in the end game 
> and that this theory applies?
> 
> - would it be interesting to add the book's problems to GNU Go tests or 
> are the situation too contrived? Or the value to low?

I have that book. I'm not sure whether the theory in that book 
will have applications or not. They transform the endgame problem
into another problem in which the aim is to make the last move.
This may work for problems where you are squabbling over a single
point but does not seem to make sense for a real-life engine. I'm
actually not completely clear about this point in their book.

On the other hand, combinatorial game theory seems to be
an area with great potential for the endgame. The basic
paradigm is that the global game is a sum of local games.
Human go players follow such a paradigm when they classify
value of each local game is typically classified as
"1 point gote", "2 points sente," etc. GNU Go makes a
similar effort with their scheme of value, followup value
and reverse followup value.

But actually these classifications are oversimplifications.
Berlekamp and Wolfe have a more scientific classification
of local games. 

There is a local problem and a global problem. The local
problem is to reduce each local game to a subtree in
which each move consists of either tenuki or making then
best local move. The global problem is to describe the
sum operation whereby local games are combined to a
global game, at least well enough to select the best
move.

Martin Mueller (who I think was Berlekamp's student at
Berkeley) is a big advocate of combinatorial game theory.

Dan