[computer-go] Seeking advice.

[David G Doshay]

Well, I am in California, not Ohio, so email interaction is about all
I can offer.

My first thought is with reference to your statement "someone ...
who knows what they are doing." I've been doing this for 3 years,
and most of the time I still do not know what I am doing. Early on
there were times I thought I knew what I was doing, but in retrospect
I can say that I only knew what I was trying to do. I find that much
of my time has gone into collecting data in a systematic way, and
what I conclude is that this game is a continuous source of mystery.
I have run numerous tests that I was sure were going to give me a
particular answer, only to find that it just was not that simple.

I strongly suggest that you also find a way to test your changes in
a systematic way. You have an advantage now that I did not have
when I was starting, and that is Don Dailey's great contribution:
cgos, the continuous 9x9 computer Go server. This is a great way
to make changes and then test them against a number of programs.

While I decided to start by wrapping a shell over GNU Go, I think
that is only one way to make progress from a known base. You
have your legal move generator already, so it makes sense to
keep moving forward from there.

In terms of overall architecture, I advise keeping the number of
free (adjustable) parameters to a minimum. The biggest problem
that I now have is that with about 20 parameters, adjusting them
in a systematic way is nearly impossible. It is so time consuming
that most parameters are just left at "first wild-guess" unless some
feature seems to just jump right out and suggest the need to fix
a clear problem. I saw this with our influence weighting, where
huge open moyos got very high numerical scores so the engine
would play them, but the resulting shapes would be full of holes.
I came to call them Swiss Cheese moyos. I cut the weighting to
0.4 of the original, and then things looked more reasonable and
we tended to loose fewer games against humans, but I still have
no idea what value would be optimal.

Other than that, I think most of us are just trying ideas that seem
like they might be interesting, and from reading this list over the
past 3 years, many of us are repeating mistakes that others have
already convinced themselves don't work.

But is has been fun and interesting, so ... welcome and enjoy.
Ask questions, because most of us are very willing to share
what we have found.

Cheers,
David



On 25, Jul 2006, at 9:54 PM, Joshua Shriver wrote:

> While lists, websites and papers are nice. You just can't beat 1 hour
> talking with someone in realtime who know what they're doing.