Re: computer-go: Computer-Human Comparisons

[Heikki Levanto <heikki@xxxxxx>]

On Mon, Dec 09, 2002 at 03:47:34PM -0000, Churchill, Julian wrote:
>  1) How much do computer programs play like human Go players?
>  1a) And how far are they designed to play like human players rather
>      than computer players?
>  2) Is it desirable to have computer Go programs playing like human players?

Difficult questions. I don't think we can have sure answers for a long time.
We need to understand better how good humans play, and how good computers
play (once we get to the point where we have good computers...)

There are a few obvious differences visible today:

  - computers tend to evaluate the whole situation at every move, humans
    play a whole game. Best humans keep an idea of the score all the time, 
    and even better amateurs evaluate the position a few times during the
    game, but most moves are made on the basis of earlier considerations.
    Humans have a focus on the important aspects of the situation, and can
    largely ignore irrelevant details. Computers are pretty weak in this.

  - Human processing happens largely in parallel, recognising patterns
    (probably even very high-level patterns). Computers need serial logic,
    a program they follow step by step. This makes it hard to analyze 
    the way humans think.

  - The same discussion happened with chess, until a powerful machine
    approach was found (brute force search). The same approach does not 
    work very well for go, but possibly some day we have tried most 
    reasonable methods, and one of them emerges as a winner. (In backgammon
    that seems to be a neural network...)
  
  - Computers can have great collections of joseki and patterns, and often
    appear to play well as long as they can follow them. But they do not 
    have much deep understanding of them, so they tend to get hopelessly
    lost if the game deviates away from them. Same happens to some humans,
    but not nearly on the same scale...
    
  - I think most go-programmers want to make the program play as well as 
    possible, beat as many opponents as possible, and gain a solid good
    ranking. Matters of playing style are (at this point?) less relevant.
    If we ever get so good programs that we can afford to loose even a tiny 
    bit of playing strength, then we can tell the program to play more
    human-like.
    

If we want to gain more understanding in how humans play, I think it would
be good to study how humans learn to play, not how the great masters play.
For they have internalised so much of the game, that they can no longer
explain how they do it...


These were my first impressions on the matter. I have been following
computer go for some years now, have started a number of small go programs
but not finished a single one. These days I hang around the Gnu Go project
instead.

-Heikki


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