[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [gnugo-devel] experimental code

From: bump
Subject: Re: [gnugo-devel] experimental code
Date: Fri, 16 Aug 2002 22:52:25 -0700

> Aside from that, I went looking at archived discussions of
> --experimental-connections, --experimental-semeai, etc.  It looks to me
> like the general consensus is that this code should be turned on by
> default, since it appears to improve things and we want bug reports, but
> that it is too slow.  So, I propose to make a level 11 of play that turns
> these on, but is otherwise equivalent to the current level 10 (I seem to
> recall that GNU Go currently gets weaker beyond 10?).  We could then make
> level 11 the default, but play online with level 10, or perhaps do
> something to check the time constraints.

At one point since 3.2 we seriously discussed turning on the semeai
code by default but some testing seemed to show it was not ready
for prime time. Gunnar ran the experimental connection code for a
while on NNGS and our rating dropped precipitously during June,
then recovered when he turned it off again.

It seems that we need to review the semeai code and try to fix
bugs but making it the default would be premature right now. I'm
not sure the level 11 scheme accomplishes anything really.

A better strategy might be to adapt the matcher_check scheme
and search for positions where a different move is made if
the semeai code is enabled than if it is turned off. This
can be tested using the gtp.

Then we could compile a larger testsuite of semeai positions


reply via email to

[Prev in Thread] Current Thread [Next in Thread]