Re: [gnugo-devel] Three stone games

[Gunnar Farneback]

Morten wrote:
> I expect the difference of one increased Handicap stone to be linear. 

Don't rely too much on that assumption. Especially not for huge
handicaps.

> I also find it very difficult to understand, why gnugo does not
> increase its strength with computing speed? If gnugo plays against
> itself with different time limits, then how many minutes will it
> need to play 1 kyu stronger ?

In principle it's possible to design a go program so that it will
asymptotically play perfectly given a huge amount of time (and
possibly memory), simply by doing a deeper and deeper global search.
This has very limited practical applicability though, because static
evaluation of the board is only trivial at the very end of the search
tree (which is later than people usually stop playing). For all
reasonable search depths, the static evaluation is extremely
difficult and moreover very sensitive to any misevaluations of group
status, connectivity, and so on.

As it happens many programs, including GNU Go, don't bother with
global search at all. It does quite a lot of local search though. The
depth limits involved in these can certainly be increased if more time
is allowed, but this only helps to a limited extent because:

1. Even with correct results of all local searches being done, the
move selection depends on many heuristics which aren't helped by more
available time.

2. The local searches are quite narrow with various heuristics to
select moves. In order to maximally benefit from huge amounts of time,
these would need to be modified as well, not only the explicit depth
limits. 

To summarize this discussion, GNU Go and probably most other go
programs as well, are not designed in a way which automatically makes
them stronger just by giving them more time or faster hardware, at
least not beyond some limit. It is certainly possible, and probably
not even hard, to make modifications which would make GNU Go play
substantially stronger if it was, say, allowed one week per move on a
fast computer. The problem is just that those changes would require
a significant amount of testing and tuning, which is simply
infeasible.

> When I started to play go, a korean 3-dan ranked player gave me 9
> handicap stones, on a 9 x 9 board. Guess who won ? I had beaten a
> korean nearly professional. And I was proud. Later he gave me 27
> Handicap stones on a 19 x 19 board. If we cannot on any board size
> with infinitely many handicap stones, like this:

It's simple to arrange e.g. 136 handicap stones so that the opponent
can't get a single alive stone even if the computer passes for the
rest of the game.

> please let a 9 dan professional try to play against this Handicap level
> ?

If you have the contacts and resources required to have a 9 dan
professional play GNU Go you're welcome. We'll be most interested in
any results.

> I believe that even only 64 Handicap stones, would firmly beat most
> drunk shodans, even in nearly total darkness, within 10 minutes, but
> that could give us the reliable feedback, that if we could achieve
> 50 % with 64 Handicap stones against 100 different 1-dan amateur
> players, each playing 10 games, then gnugo has finally reached the
> level of 64 Kyu. Which is my assumption.
>
> If this is the true, we have to confess it. 
> I believe it is a reliable way to find out the real-life strength.

Not at all. I don't think it's hard to tune GNU Go to make it
virtually unbeatable with some 30 or so handicap stones, but it would
have absolutely no bearing on its playing ability in even games or
with reasonable handicaps. And frankly, it would be a total waste of
time.

> And then we could try to increase the strength with decreasing 1
> Handicap stone, every 3-6 month or so.

If you think this is a good strategy to get a strong program, I'm
afraid you don't know what you're talking about.

> It is no real achievement to play against ManyFaces or Handtalk, 
> when they perform moves lightyears apart from go moves played by
> experienced dan ranked go players.
> 
> We do want to beat a professional human, not ManyFaces or Handtalk ?

In the short term it's very interesting to beat Many Faces or
Handtalk, yes. To reach the top of the go programs would be a major
achievement with significant publicity value. And it's infinitely more
interesting than to beat professionals with 30 handicap stones.

Notice that we also do lots of tuning on games against humans, mostly
ones played on NNGS. There GNU Go is rated 13 kyu, which I think is a
quite reasonable estimate of its real-life strength.

/Gunnar