Re: [gnugo-devel] 261 games 3.1.31 vs 3.1.32

[Trevor Morris]

At 12:05 PM 4/19/2002 -0700, address@hidden wrote:
>> By strange, I was thinking about cases where both GNU Go instances seem 
>> to avoid contacts and fights and make many extensions (ikken tobi, 
>> kogeima, kosumi...). This sometimes result, after 50 or so moves, in a 
>> large spread of "loose" stones all over the board, with patterns that 
>> seem unnatural to me, or at least very different from what I see in 
>> human games. Am I making sense? And if so, this is probably to be 
>> expected, but then I would think that it shows that playing GNU Go 
>> against itself is not a very good measure of its progress or strength.
>
>I agree that although GNU Go 3.1.32 can give GNU Go 3.0 a 3 stone handicap it
>is not 3 stones stronger against humans. On the other hand it has not been
>explicitly tuned against 3.0, so the benchmark is not completely
>meaningless. I think the increase in strength over 3.0 is about 2 stones based
>on results on NNGS.

If GNU Go can continue to increase at 3 stones (against itself) every 6-9 
months,
I believe that it will get quite strong in a few years.  Even if the tuning 
were to
previous versions of GNU Go (which it's typically not, and probably won't be),
this is not an insignificant improvement.  It's also the simplest and most 
objective metric we have to measure performance increases.

In short, I think that using older versions of GNU Go to test the strength of
newer ones is valid.  It will be interesting to see how many stones future
GNU Go's will give to 3.0 and 3.2 respectively.

-Trevor