Is your reference compiled java version on cgos?  I'd like to cut back Many
Faces to your spec and see how it does.  Maybe find some bugs.

> -----Original Message-----
> From: [EMAIL PROTECTED] [mailto:computer-go-
> [EMAIL PROTECTED] On Behalf Of Don Dailey
> Sent: Saturday, October 18, 2008 6:26 AM
> To: computer-go
> Subject: [computer-go] reference bots testing.
> 
> I have two versions of the reference bot.  A C and a Java version.
> 
> I am doing a conformance self-test on my java and C versions and they
> do
> not agree.   The simple stats seem to agree but they play at different
> strengths when self-testing with actual games.   The C version is
> playing about 40 ELO weaker and it's less and less likely to be
> statistical noise.
> 
> So I found a minor bug, fixed it, and run the tests again and still
> find an error, but it looks to be around 25 ELO after a few thousand
> games.
> There are enough games that this bias looks consistent.
> 
> The weaker C version, uses the Mersenne Twister which I have some
> confidence in, so I really doubt the RNG is a factor here since the C
> program probably has the superior RNG.   I am also careful that I
> select
> numbers in a uniformly random way - a lot of people do this wrong
> because  stuff like (rng() % N) doesn't do what people think it does.
> 
> So I have a little mystery here.  The program is pretty straightforward
> and simple, so it proves something that the chess people know,  that
> the strength of your program is heavily dependent on how many bugs you
> have and that straightforward bug-free programs tend to be pretty
> strong.  It has been said that the secret of Fruit, a very strong chess
> program is mostly that it is bug-free, there is nothing spectacular
> about how it
> was written other than this.   I don't know to what extent that is true
> for fruit,  it's just something that has been spread around,  but I
> know that it tends to be true for me and my software.
> 
> Even if I find the bug(s) and fix it,  I am still comparing two of my
> own implementations.   So it would be good if someone else would send
> me
> their "conforming" version (for linux preferably, although I can
> usually run exe files) and of course my java implementation is freely
> available.
> 
> I could also send a compiled java version (that does check out against
> the JIT version), a binary for 64 bit linux.   That version is actually
> faster than the latest IBM java JIT by a good little bit and it has a
> smaller memory footprint, faster startup time, etc.  It almost makes
> Java a practical language.
> 
> - Don

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

Reply via email to