[gnugo-devel] Re: zergling

[zergling]

I do not reply quickly because It is a little hard to express what I think.
My english is a little poor. So I have to look up dictionary frequently when
post something. It really take much time. I agree that my idea is not
mature.


I just wonder why there are many post about the detail implemations but few
about ,
I dont know how to express it , "the main point". And very few discussions.
Perhaps you should think more about the second , if gnogo's goal is not only
a "gnu" go.

There are no key difference between the algorithms of mfgo and gnugo. GnuGo
is stronger now.
You correctmany bugs ,add many patterns and tune the porformance. But the
architecture is same.
I think mfgo is mature enough, but it still play like an amateur.

Then I continue to post my opinion, I hope it is not valueless.

The ideal go program: efficient search algorithm, an evaluation function can
descibe the concept such as
territory and influence, machine learning mechanism.

Game like go, chess etc are search based. There is a game-play theory based
of search.
But there are some differece between go and chess. Chess has little
knowledge except the rules.
People play chess is mainly based of search. But if u want to play go well,
u should
get many knowledge. But this knowledge is generated from search. Because at
the first
there are nothing except the rules and game-play theory. So I am a little
surprised why
gnugo use search to judge connection, death and life etc, but has not a
search in the main body.
It is not global search, it is just a search base of the result of local
search.

And we know search is expensive. So how to make search efficent is the key
to improve
the strength of go program. Global search is hopeless. But go is local game.
Every move
just affect locally. Thinking of this architecture: hiberarchy search, a
high level search is based
on the low level search. High level search  such as global search has large
scope , but low level search
such as search to judge connections has small scope. And this is the way of
people to play go. There
are a good paper to descripe a plan-based search. It split the main plan to
sub plan. URL
http://liawww.epfl.ch/~willmott/Edinburgh/snw.ps.gz. I think it is similar
with my idea.

Any search's result is conditional. if the surrounding change , result may
change.
A search tree has a set of positions to limit it. The positions set is made
up of the branchs's position set.
If there is stone at the set.  It should refresh some affected branchs. So
every move  only affect
a little trees. If we record the tree, it will evidently accelerate the
search. Perhaps it eat many memory,
But it is worthy.

Definit that the tree's scope is the limit positions set and the branch's
node's position set.
But If we find two tree's scope is overlapped. Then we examine the
overlapped scope to find
good move.

And if a local tree apear frequently, we can record it to accelerate search
.

The knowledge except the rules is generated from search. I agree that there
should be some databases to store
knowledge. But it is important for computer not only to use it but to
understand it. The professional player
just remember some common joseki. If forget a joseki, he can easily get it
by search and some other knowledge.





----- Original Message -----
From: "Manner Wolfgang" <address@hidden>
Newsgroups: gmane.comp.games.go.devel
Cc: <address@hidden>
Sent: Monday, March 25, 2002 9:49 PM
Subject: zergling


> i think we should welcome provocative views. ideas which do not
> conform to the general party line could very well be the right direction.
> of course it is not enough to say: "you are stupid" (for myself i would
> accept that right away) you also have to have something to say.
> anyway handtalk has some very interesting features, it would be good
> to know, how it works and maybe one could profit ?
> so zergling please continue to critisize !
> wolfgang