[gnugo-devel] dfa_unsort_1_25.1

[Arend Bayer]

In matchpat, there is a compiler switch DFA_SORT. I assume this was mainly
introduced for debugging purposes, to make dfa behave as close as possible
to the classical pattern matcher. (The patterns get sorted in the way
they appear in the database before the owl callback is involved.)

However, it is clearly redundant, as the owl callback function immediately
sorts them again. Also, it consumes 4.2% of CPU time. So if we trust the
DFA, we should probably take this down.

I ran full regressions without DFA sort. An analysis of the results is
below. Altogether, there are 3 FAILs. ld_owl:301 is clear-cut and important
(and one should try to repair this by an epsilon correction of pattern values I
suppose); with nngs:1660 I am not sure whether the changed owl opinion is
correct; the FAIL in global:4 is caused by a seemingly correct change in
owl results.

I don't know whether these results should imply that we should make this
change immediately, or first see whether we can get ld_owl:301 right.

Arend

./regress.sh . ld_owl.tst
301 unexpected FAIL: Correct '3 B18', got '0'
This is a serious mistake.

./regress.sh . trevorb.tst
920 unexpected PASS!
This is apparently a platform dependency.
Further, the move played is is not an improvement (GNU Go plays A12 instead
of B11), problem needs revision.

./regress.sh . nngs.tst
1660 unexpected FAIL: Correct '!F18', got 'F18'
Status of R14 changed from critical to alive. I have no idea whether this
is correct.
1690 unexpected PASS!
Another platform dependency? (I get the same result with 3.1.24.)

./regress.sh . trevorc.tst
280 unexpected PASS!
Yet another. My 3.1.24 claims that C9 attacks the worm B7 with good ko. This
is incorrect.

./regress.sh . global.tst
4 unexpected FAIL: Correct 'Q6', got 'F5'
The valuation of Q6 is not changed. However, with dfa_unsort GNU Go thinks
that F5 kills E4, which is correct I believe.
The valuation comparing Q6 and F5 is then a little bit wrong. (Although I
think it is a close decision.)

./regress.sh . arend.tst
29 unexpected PASS!
This one is strange. I cannot reproduce this from command line nor with a
direct gtp command.