Re: [gnugo-devel] A serious scoring error

[Gunnar Farneback]

Stephane wrote:
> Game 45: W+14.5 W+20.5
> 
> So White and Black disagree by 6 points. Moreover, when I replay
> the attached game, I find that Black wins by 19.5 if we assume
> that the white group on the right is dead (komi was 6.5).
> 
> What should the exact score be ? Could somebody familiar with the
> scoring code have a look at it ?

B+19.5 is indeed the correct score. To see what goes wrong, the
standard trick is to run

gnugo --quiet -l test_match_1_045.sgf --score aftermath -o x.sgf

and look at the output sgf file. Apparently something strange goes on
at move 7 (white R6) and 8 (black P7).

We can also see that

gnugo --quiet -l test_match_1_045.sgf --score finish

correctly says B+19.5 so the problem is limited to the aftermath
scoring.

So far I'd guess that the mistake is either caused by a bug in the
aftermath code or a serious owl mistake. To get further we need some
traces, like

gnugo --quiet -l x.sgf -L 7 -w -t -d0x1800

We can see that R6 is supposed to be an owl defense for S10, which is
bogus. Looking closer at the owl reading of S10 at the next move,

gnugo --quiet -l x.sgf -L 8 --decide-owl S10 -o y.sgf

and analyzing the output sgf file the problem appears to be a rather
severe escape potential misunderstanding around variation 160 (after
B:S6, W:R8, B:Q6). Escaping towards the tactically critical N8 stones
is supposed to be useful, which clearly is nonsensical. It's worth
noticing that this mistake didn't appear in the original position
since the N8 stones were connected to the S10 dragon then, providing
no illusory escape potential.

The appended patch adds this mistake to owl1.tst.

Thanks for pointing out the problem.

/Gunnar

Index: regression/owl1.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/owl1.tst,v
retrieving revision 1.68
diff -u -r1.68 owl1.tst
--- regression/owl1.tst 10 Jun 2003 04:59:49 -0000      1.68
+++ regression/owl1.tst 6 Jul 2003 11:02:39 -0000
@@ -238,6 +238,11 @@
 318 owl_does_defend F10 D10
 #? [0]*
 
+# Problem with bogus escape potential in GNU Go 3.3.22.
+loadsgf games/owl47.sgf
+319 owl_attack S10
+#? [1 S6]*
+
 
 ########### end of tests #####################
 
Index: regression/games/owl47.sgf
===================================================================
RCS file: regression/games/owl47.sgf
diff -N regression/games/owl47.sgf
--- /dev/null   1 Jan 1970 00:00:00 -0000
+++ regression/games/owl47.sgf  6 Jul 2003 11:02:41 -0000
@@ -0,0 +1,20 @@
+(;GM[1]FF[4]
+SZ[19]
+GN[GNU Go 3.3.22 load and print]
+DT[2003-07-06]
+KM[7.5]AP[GNU Go:3.3.22]RU[Japanese]
+;AW[ia][hb][jb][nb][ob][pb][ac][bc][cc][dc][jc][mc][nc][oc][pc]
+[dd][fd][jd][ld][md][nd][qd][ee][je][ke][ne][af][ef][nf][of][eg]
+[fg][gg][hg][jg][kg][ng][hh][ih][kh][lh][mh][ei][li][si][ej][fj]
+[jj][rj][dk][gk][hk][jk][kk][qk][rk][dl][fl][gl][hl][il][jl][kl]
+[ml][pl][dm][fm][hm][lm][mm][nm][pm][rm][dn][fn][nn][qn][eo][fo]
+[no][ap][cp][dp][fp][gp][jp][aq][bq][cq][fq][hq][iq][jq][nq][pq]
+[qq][rq][ir][kr][nr][or][pr][rr][sr][js][ks][ls][ns][rs]
+AB[ja][ka][ma][na][oa][pa][kb][lb][mb][qb][kc][lc][qc][ad][bd]
+[cd][kd][od][pd][rd][ae][ce][de][le][oe][pe][bf][df][if][pf][bg]
+[dg][ig][og][dh][eh][fh][gh][jh][nh][ph][rh][di][fi][hi][ii][ji]
+[ki][mi][ri][cj][dj][gj][hj][ij][kj][lj][nj][pj][qj][ck][ik][lk]
+[mk][pk][cl][ll][cm][gm][im][jm][km][cn][gn][hn][ln][mn][on][ao]
+[bo][co][do][go][io][jo][ko][mo][oo][ro][bp][hp][ip][kp][lp][np]
+[op][pp][qp][rp][sp][kq][lq][mq][oq][sq][lr][mr][ms]
+PL[B])