Re: [computer-go] Minimax with random evaluations

[Mark Boon <tesujisoftware@xxxxxxxxx>]

On Mon, 24 Jan 2005 08:11:45 -0500, Don Dailey <drd@xxxxxxxxxxxxxxxxx> wrote:
> 
> >  In many games having lots of move options correlates somewhat with
> >  a good position. But, as noted previously, this is not very
> >  relevant in the context of go.
> 
> It looks like it may be relevant in GO too.  Of course this depends on
> what you mean by relevant.
> 
> Random evaluation is a very poor evaluation in any reasonable game you
> choose, but the interesting result is that random evaluation combined
> with a search leads to something much better than completely aimless
> play.
> 
> In my tests of 5x5 Go, 5 ply random evaluation beats random play with
> an 81% score over 1000 games.
> 
> In 7x7 Go, 5 ply random evaluation beats random play with an 82.4%
> score over 523 games.
> 
> Not very impress when you consider the huge amount of extra work being
> done, but quite impressive when you consider the evaluation function
> used!
> 

I think these findings are very intriguing. But how about putting this
a bit more to the test, and compare '5 ply random' against '5 ply
legal-move-count'? That may remove the doubt from your "It looks like
it may be" statement above. If your assumption is correct, the latter
should have a 50+ winning percentage, possibly also up to 80%.

As to the relevance: when I have more time I may write something more
elaborate on this, but I think from previous posts here it's obvious
the relevance is fairly high. There have been some arguments that more
reading didn't improve the level of play due to the fact the
evaluation function was too unreliable. Your experiment may give a
better boundary to how unreliable it needs to be to get worse play
when reading deeper. Or it may show that that argument has been
completely false all the time, and the reason has to be looked for
elsewhere.
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/