On Sat, 2008-10-11 at 13:34 -0500, Zach Wegner wrote:
> >  7.  In the case of moves with even scores a random selection is
> made.
> 
> I think maybe this should be something deterministic.

That of course is clearly a possibility.   However I'm trying to
approach this with a certain consistent methodology which is to specify
the BEHAVIOR, not the implementation.   I want complete freedom in how
you implement the specified behavior.  


Also, if you chose a different generator, it would not be possible to
prove it if it were a good one.   Not unless we also got heavy handed
about how it's initialized, and how you select moves with a random
distribution.    In short, I fear that we would be asking too much to
build a bunch of clones that are completely deterministic.

If you look at pseudo random number generators you will see there is
this same implicit methodology, though not stated.   The behavior
specified is to "appear random" and test are designed to see how well
this behavior is simulated.    In some sense, people consider the best
random number generators to be the ones that "fool" the Diehard tests.
They fool these diehard tests into thinking they are really random. 

I want this to work that way.  We will build tests and I don't care if
you can find a way to fool the tests.  However the tester will be
designed to be difficult to fool and it will not present you with
predictable positions.  

I don't care if you don't follow the spec 100% if you can fool the
tests.   If you can do this, you will probably have found something
useful to all of us. 

This is not going to be a formal thing like the Computer Benchmarks Game
is.   We will just build a tester and publish the spec and you can do
what you want with it.   Maybe put entries on Sensei's page.  

- Don

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to