[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: computer-go: Computer Go hardware



On Sat, Mar 31, 2001 at 12:22:56PM -0500, Compgo123@xxxxxxxxxxxxxxxxx wrote:
> No matter how you look at it, a significantly improved hardware is essential 
> for the computer playing to reach a high level (temperorily disregard the 
> fact that the jury for a magic formula is still out). I don't want to argue 
> whether an improved hardware will help. For those who claim it doesn't, 
> please point out what will.

"No matter how you look at it, significantly improved vitamins are essential
for eternal youth. Temporarily ignore the fact that we have no idea of how
to gain eternal youth. For those who don't believe in the importance of
vitamins, please point out what gives eternal youth!"

Hardware improvements are good *after* we have solved the theoretical
problems - or at least isolated the questions we need to answer. As things
are now, most theories can be tested and developed on today's hardware, and
only of we find something that plays well, but too slow, can we consider
speeding it up.

> PC cpu is now running at about 1 GHz. Intel is planning to introduce a 10 GHz 
> cpu in next 5 to 10 years. 

That is only a factor of 10. If you were optimizing a large bubble sort, a
factor of 10 would be good - until someone came up with a quicksort. Also in
Go, I believe we need - and can achieve - algorithmic improvements of such
magnitudes that ther relatively slow advances in computing power pale in
comparison.

> We may define a concept, calling it the 'logic efficiency index'. 

You may define as many concepts as you like, but I feel that if they should
have some effect on computer go, the most effective are to define them in
would be go theory, not some vague perfectionism of non-existing programs.


The way I see it, there are a few promising routes for computer go:

1)  Improved understanding of the game itself, possibly some new
representation or idea that will solve some part of the problem that
currently takes lots of hard work. Benson's theories of life, Berlekamp's
endgame analysis, Wilcox' sector lines, etc.)

2) Incremental tuning and tweaking of existing programs will keep improving
their performance, possibly a lot.

3) Higher-level planning, and other AI ideas. I believe we need to get away
from looking at the board as 361 independent units, and think on a much
higher level. Look at a typical way a human describes a situation: "Black
has a few weaknesses, but is ahead in territory". That is the *starting
point* for planning future action. Most programs do not have anything like
this level of understanding - yet.

4) Advances in self-learning systems may - or may not - bring a breakthrough
to computer go as well. Probably not a program that will learn, on its own,
perfect play from the basic principles, but a small module that can be
painstakinly trained to answer some simple questions, for example which of
two alternative positions have better chances of escaping into the center.


I am sure there are other important advances coming along, and I would
welcome discussion on them, even if they relate to hardware...


- Heikki

-- 
Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>