computer-go: Sucessful Computer Go Characterization

[Matt Gokey <mgokey@xxxxxxxxxx>]

On Fri, 1 Sep 2000 Heikki wrote:
>> I can well imagine a program that will analyze the situation carefully
>> (including, of course, some tactical reading about life and connections),
>> describe it in some higher-level language ("weak group here", etc), make
>> plans on that level ("force the weak group to run this way, and secure
>> territory while attacking it"), and then find a move that implements the
>> plan (again using local tactical reading where necessary). If any (global)
>> searching is to be done, it will be done on a high level, among relevant
>> plans, (almost) disconnected from the board. I would like to experiment with
>> something like this, if I had more time.

I agree completely.  I wrote about an idea similar to this in May of
1998 after I just learned about computer go - I called in "chunking
search" in a message titled "High level ideas...".

On Thu, 25 May 2000 Thore Graepel wrote:
>> I thought you might be interested in this because we use a new approach of
>> converting a graph-based representation into feature vectors that enable us
>> to learn some aspects of good and bad "shape". These methods are then used to
>> learn good Tsume-Go moves and to play on a 9x9 board after learning from game
>> records.

A potential form of representation that may certainly pay off...

On Sat, 2 Sep 2000 Nicol N. Schraudolph wrote:
>> The central problem in any computer go program seems to me one of represen-
>> tation - if we knew how to encode the high-level concepts well, we'd be a
>> lot further along.  Good representation should make generalization easy.

I know if I had the time, this is where I would be concentrating my
efforts.

------

IMHO the representation problem and specifically multiple level based
representation and understanding of the game is the key to computer go.
I don't believe the best go programs will come from the iterative
deepening search camp no matter how complex it becomes, nor will it come
from the pattern recognition human copy-cat camp.

The end result will play much different than computer chess programs do
or than humans do, will take advantage of the computer's unique
capabilities rather than emulating humans, and rival even the best human
players.  It will be a combination of partitioning, pattern matching,
global searching, planning, and tactical searches - at various levels of
understanding/representation -
some abstract, and make use of a huge dynamic database of learned
patterns, openings, shapes, and
sequences.  It will probably use genetic programming or neural net
components or other AI methods as well as traditional searching and
evaluation techniques.  It will also likely make use of mathematical and
theoretical Go knowledge and derive everything from first principles,
rather than being hand crafted by a expert player with hard coded
heuristics.

And I think it is doable within 10 years.

Do others believe multiple level representation is worthy of serious
consideration?

What forms of representation (previously researched or new ideas) might
we use to abstract Go concepts?

How would others here characterize a hypothetically successful Go
program?

Matt