Re: computer-go: Insight of a human continued

[Heikki Levanto <heikki@xxxxxx>]

William Harold Newman <william.newman@xxxxxxxxxxxxxxxxx> wrote:
[how to evaluate tenuki]

> Not to criticize the way Goliath does it, which is probably quite
> effective in most cases, but people should realize that getting this
> right in *all* cases is a nontrivial problem.

Of course it is. Playing Go perfectly is something only the best of humans
can approach, after a lifetime of studying.  This is a far cry from the
state of modern Go programs... I believe the Go programs can be improved
upon by simple heuristics and programming tricks. The day when we start to
approach the limits of theoretically possible, are very far ahead indeed.



>   * The value of a position is no longer simply a real number,
>     but an element of a nontrivial group, with only partial
>     ordering. 

This is known in "classical" Go theory as well, at least to the point that
sente and gote has to be evaluated in addition to the simple value of a
move. I am quite willing to believe that an exact mathematical formulation
of sente and gote will have to go several levels deeper, and in the worst
case consider the game all the way to its end before it can come up with a
precise answer.


>   * Even once you have the values of a set of positions, it is
>     in principle NP-hard to determine who wins. But note:
>     ** In practice, this NP-hardness might not be a significant
>        problem, since various approximate analyses seem to be
>        very effective in real games.


Yes, indeed. After all, we want us (and our computers) to play pretty good
in real games, not in some theoretical constructs.

Sorry if I sound like racking down on these theories. I do think them to be
valuable in their own right. I just don't see the current computer-go
situation benefiting much of them.

- Heikki

-- 
Heikki Levanto     LSD Levanto Software Development   heikki@xxxxxxxxxxxxxxxxx
               "In Murphy we Turst"