On Thu, Jun 3, 2010 at 6:25 AM, Scott Christensen <[email protected]>wrote:
> I agree that the human brain is better wired to play Go than Chess.
> Humans can remember hundreds of thousands of visual images and their
> significance, but we only have a working memory of 6-7 steps in a
> temporal sequence. There are fewer possible moves in chess and its
> easier to assign point values to positions, so a computer can make
> moves based on expected consequences beyond 6-7 moves in the future.
>
> In Go there are too many possible future combinations for even
> computer systems to deal with. Humans rely on pattern recognition for
> effective game play. Computers are extremely poor at pattern
> recognition of salient features and almost totally lacking in
> intermediate goal setting which are extremely important to human play.
>
> Computers are not really 'thinking' but are merely sorting data to
> come up with a 'simulation' of a proper game move. There are many
> board scenarios that can be presented to game systems that demonstrate
> they have absolutely no concept of the games.
>
> The newly popular technique of Monte Carlo Tree Search does go a step
> closer to a human thought process of predicting the consequences of
> moves to a distant future outcome rather than just calculating point
> values a few moves ahead which is perhaps why this technique has had
> such fantastic success lately.
>
>
I think your assessment is accurate. Of course chess suffers all the same
effects you mention for go, just in different proportions. I think that
fact is not understood. Even the very best programs can search 20 ply
ahead and not understand what is really going on in computer chess -
although with increasing depth that is getting rarer and rarer. And there
are positions that can be presented to chess program that reveal they are
essentially clueless. The amazing depths that computers search now
covers a multitude of sins, but the same problem is still there, the
evaluation function is completely broken for some positions. It only
works if we can search deeply enough that we are no longer evaluating the
trouble positions.
The amazing progress in computer go is based on the same exact paradigm that
chess uses, evaluation and search. Since the games have different
characteristics we do each differently, but that is just semantics - it is
fundamentally search and evaluation. Intelligent playouts turn out in GO
to be far superior as an evaluation function than the old static methods
and depth first search is replaced by best first search.
And by the way, even old style brute force alpha/beta "works", it's just
that there are better ways by far so it's rather silly to use this directly.
This is also the way it is in computer chess - a program that does not
prune still gets much stronger with depth (so it "works") but we have gone
way past that.
Don
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
