[gnugo-devel] Good semeai results!

[Daniel Bump]

The last change to the experimental semeai module was a change
that went directly and accidentally into the CVS on Christmas
eve without a patch recorded or ChangeLog entry. (Of course
this isn't supposed to happen.) This patch is:

http://savannah.gnu.org/cgi-bin/viewcvs/gnugo/gnugo/engine/semeai.c.diff?r1=1.28&r2=1.29

What this patch does is turn off the semeai module for
a large class of examples that were causing problems.

(The subsequent patch gunnar_1_20.9 affects the
standard semeai code, not the experimental code.)

Apparently this patch was a good one because the experimental
semeai code is now working much better. Compiling 3.1.21 with
'configure --enable-experimental-semeai' 
gives 13 unexpected PASS's and 3 unexpected FAIL's.

The 13 unexpected PASS:

strategy2:72
nicklas1:1406
nicklas3:602
nicklas5.tst:1204
trevor:461
nngs:1600
nngs:1700
nngs:1709
global:46
trevorb:430
trevorb:630
trevorc:800
strategy4:188

The 3 unexpected FAIL:

arend:31
13x13:25
strategy4:168

Presumably these PASS's were there before and what
the patch accomplished was to eliminate a lot of bad
results from positions that were not good semeais.

I'm not sure how the semeai version is in terms of speed.
We could consider implementing Zobrist hashing for it.
It also needs to be taught about kos.

I am currently running the regresions with --enable-dfa
(an no semeai). I'll report on those results tomorrow.
At this moment, there don't seem to be many difference
between those results and those without the DFA. This
is good news, and is in contrast with the situation a few
releases ago, though I have no idea what change could
account for it.

A couple of months ago Teun proposed that we make the 
DFA compilation standard. Perhaps he is right.

Dan