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

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 20:47 1-12-2004 +0100, Marco Scheurer wrote:
>On Dec 1, 2004, at 20:45, Vincent Diepeveen wrote:
>
>> Of course that doesn't require a 30 ply search in itself. Probably 15 
>> ply
>> is more than sufficient.
>
>You still need to have a decent evaluation fonction, don't you?
>marco

A good agressive evaluation function is the first necessity to solve *all*
problems. 

I remember how my draughtsprogram i could let search 30 ply overnight in
openingspositions (just a few moves played) and it still would play the
utmost beginners move. 

Algorithms like nullmove even exponentially strengthen the effect of
evaluation functions.

Now writing a good go-evaluation function for 19x19 isn't easy.
Additionally it will search quite slow too.

So in that respect a HydraGo will amaze you all as i know what he'll do to
build it, if Chrilly would get a budget for it (i'm sure something like
2000 euro a month would more than do for such a project).

I am guessing i know what todays 19x19 programs are doing and that's of
course very sophisticated, but i also realize very well the life&death they
are missing because of how they search. In hardware that all goes quite a
bit faster.

In chess, where full board evaluation takes considerable less cycles than
go, the trade off between fast search with primitive agressive evaluation
function was quite unclear until recently and when something like that gets
unleased in go (primitive evaluation function compared to the better
evaluation functions) will get incredible anti positional games on the
board and win many of them because the anti positional agressive tactical
mess the tactical deeper searching for the coming few years in go has a
major advantage.

If you just attack a lot against todays commercial software you win all
life&death hands down. Professional go players could learn from this
principle a lot too! 

Vincent

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