Re: [computer-go] Pattern matching - example play

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 18:00 1-12-2004 +0100, Erik van der Werf wrote:
>Frank de Groot wrote:
>> What makes you think you can enlighten me about these issues?
>
>Your posts.

:)

>
>> The only interesting thing is to notice a statistical advantage on 9x9, 
>> that should be possible.
>
>If for example optimal play is a deep narrow line there is a very good 
>chance that such statistics derived from weak players are completely 
>meaningless.

Automatic derived statistics in general are useless, because if you play
the average move of the average player on the planet you won't *ever* make
it into dan levels, unless you back it up with a 30 ply search or so.

>> I don't consider 9x9 Go to be Go :)
>> It's another game. I am not interested in it as I think it's easier than
>> chess.
>
>
>So far, the chess programmers have failed to make a decent 9x9 Go program.

9x9 go is relatively seen to the number of possibilities on the board,
harder than 19x19 go.

Again *relative*. 

In 19x19 go you can forward prune certain moves with like 99.999% certainty.

In 9x9 you *must* consider every move.

>Personally I think that at the current level of computer Go 9x9 is very 
>interesting. After all beginners start on 9x9, not on 19x19. If we 
>cannot even build a decent 9x9 player how can we ever hope to succeed on 
>19x19?

'ever succeed on 19x19'. This is of course only a matter of time. Of course
we can disagree upon how much time. 

However there are some huge differences between 9x9 and 19x19. First of all
if some top chess programmers build a 9x9 go program in combination with a
deep search, parallel hardware, i really doubt any human will be able to
beat it.

There simply has not been put any big commercial effort yet into 9x9 go.

Majority of search experts are not capable of speaking either Japanese,
Korean or Chinese. Those markets are very difficult to position a product
into.

9x9 go really is different there. The branching factor for a selective
searching program is very similar to a chessprogram and tactics in go are
easier to extend than in chess.

I hope you realize chess software sees also forced lines up to 128 ply deep
without problems (well mine is at least).

So the basic problem for 9x9 go software will be the first 20 moves or so,
after that it gets a lot easier as search will dominate it.

In 19x19 that's at this point completely impossible for several reasons. 
  a) bigger branching factor 
  b) tactical lines you need to search deeper.

So there is 2 type of explosions from which A really is the big killer.

If some rich go player would put a big price onto a 9x9 tournament, which
is far above the return ticket prices to Japan, China or Korea from Europe,
say 10000 dollar for the tournament and a chance to play a strong go player,
then 9x9 go will have sooner a program playing better than mankind than
chess will have.

The reason for this is dead simple.

Chess has 10^44 possibilities. This has been a real difficult calculation
of someone who promoted onto that subject. Read in ICGA the details.

Go 9x9 even a simple estimate it has :
  say 25 white, 25 black, 31 empty :

less than : 81! / 25! 25! 31! = 2,9 * 10^36

Even the most idiotic approach possible : 3^81 = 4,4 * 10^38

Now reality is that when there is like 40 open fields left, many are not
very clever to lay a stone in. 

So selective searching you can search magnitudes deeper than in chess.

In chess making a function that selects the best move in a static way is
real real difficult and chances you prune the best move is a real
possibility.   

So if you compare 9x9 to chess and would put some commercial effort into
9x9 go, it's real trivial to play this at a level far beyond any human can.

Of course it will take some years to manage that, and as the years pass by,
hundreds of people at the same time keep programming chess engines and
there is only a few 'academics' with go 9x9 engines.

The big 2 borders to make a real good 9x9 go engine are:  
  a) even the x-dan players here in Netherlands are real beginners from
positional viewpoint. Not capable of explaining patterns from which i'm
sure 5 and 6 year old get them explained in Asia.

  b) the commercial selling prospects of such engine are real complicated. 

In the end it is always B which is driving developments. If several can
make a living out of it, then there will be software for it. Right now it
is what we call in netherlands with a very self explaining statement: 
  "One eyed in the land of the blind"
All those professional 19x19 engines are so optimized to 19x19 that their
entire search and evaluation would need a complete rewrite for 9x9.

Again in contradiction to what some write here, making real good searching
software is NOT easy nor trivial. Without 10 years experience there, forget
even trying it.

So principally i agree more with Frank de Groot that 9x9 go is a completely
different game from 19x19.

Calling it a kids game i don't find a real good idea. tic-tac-toe and
4-connect is what i call kids games. A human can play it perfect...

Vincent

>E.
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/