Re: [gnugo-devel] Fuseki strategy discussion

[David G Doshay]

On Jun 10, 2004, at 11:12 PM, Heikki Levanto wrote:

> Would this be a good place for a limited whole-board lookahead? I have
> an idea that in the fuseki stage, we could consider a very limited set
> of moves, to sort out a suitable direction of play. In this context, a
> "move" could be a whole joseki (the most typical variation), or some
> other typical sequence...
>
> Still, as long as our influence evaluation gives bad ideas of the value
> of the resulting positions, there isn't much use for this.

We have been experimenting with whole board lookahead and are
"almost" ready to show you all what we have done.

Indeed, culling down to a limited number of variations is the easy
part and the evaluation functions are the key. But there is another
problem that I think is equally as problematic: for the lookahead
sequences there is little choice other than to use the GNU-Go engine
as the move generator for the opponent's move. The result we are
getting so far is a great winning percentage against standard GNU-GO,
leading to the assumption that we made the program much stronger.
However, the reality is that we are only stronger against an opponent
whose moves we can predict perfectly.

In our project we are now beating back a set of bad moves that our
modified code makes but that GNU-Go does not, and then we will
start playing against other Go programs to see how we do. Reports
will follow.

Cheers,
David