Re: [computer-go] Minimax with random evaluations

[Mark Boon <tesuji@xxxxxxxxx>]

Don Dailey wrote:

> >   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'? 
> >    
> What is 5 ply legal-move-count?  Are you suggesting that I use number
> of legal moves as an evaluation function, say the difference between
> black and white move counts?  I could do this,  but no matter what the
> result I don't see how it would help resolve anything in my mind.
>  
Yes, that is what I suggest.

> What would you conclude if say ...
>
>    1) move count did much better.
>  
That the reason random evaluation is improved by reading is in fact only 
caused by choosing lines of play leading to bigger move counts.

>    2) move count did much worse?
>  
That this is not the case, and another explanation has to be found to 
why reading helps in case of a random evaluation. I can't see any other 
reason so that would make me very curious.

> My intuition is that in chess, move count would be much stronger than 
> random evaluation because it is a more or less direction measurement
> of mobility,  a very important concept.   
> I also believe move-counts, if it's what I think you are saying, would
> be a lot better in GO than random evaluation.  It would cause a
> program to want to make a lot of 1 point eyes (because they are
> illegal for the opponent to move into, thus the computer would be
> happy to keep a point alive that only it could move into.  This of
> course assumes suicide is illegal.
>
> I'm thinking random evaluation has a similar effect and may be the
> main benefit, but not to the same extent as move-count mobility.  Just
> a guess, I don't really know.
>
>  
I'm surprised 5 ply reading leads to such a big winning percentage in 
case of random evaluation. The only explanation I can think of this 
happening is legal move-count. If one side has many more moves, then 
there's a bigger chance the random number generator will generate a 
large number among them. So it tends to favour lines of play that have 
more legal moves. I say 'tends' because by nature of the random number 
generator, many moves still do not guarantee a high random value, it's 
just more likely. So by this reasoning, an algorithm that removes this 
uncertainty by directly counting the legal moves instead of indirectly 
by generating random numbers, should be better still.

This is an assumption I'd like to see tested. I share your intuition 
that move-count matters, also in relation to Go, especially on a small 
board. I expect it to make less difference for bigger boards. But 
intuition can have the awful characteristic of being completely wrong, 
and it seems there's an easy way to test this hypothesis.

   Mark

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