playing into snapback

[gunnar]

In  http://lists.gnu.org/archive/html/gnugo-devel/2006-01/msg00033.html Dan wrote:

Recently I've seen GNU Go take snapbacks on a number
of occasions.

Here's an example. Current CVS plays W at Q7. This
is a contrived example but this sort of thing does
happen in actual play.

Move generation values P2 to 9.24
  Q7: owl attack/defend for R6
  Q7: owl attack/defend for S8
  Q7: owl attack/defend for S6
  Q7:   2.33 (followup) - threatens to capture Q6
  Q7: not a safe move
  Q7: 17.01 - change in territory
  Q7:   7.20 - S8 strategic value already counted - A.
  Q7:   2.80 - R6 strategic value already counted - A.
  Q7:   3.60 - S6 strategic value already counted - A.
  Q7: 3.17 - strategic effect on Q5
  Q7:   13.40 - total followup value, added 11.07 as territorial followup
  Q7: 6.70 - added due to followup (13.40) and reverse followup values (0.00)
  Q7: 1.24 - shape (shape values +1.00(1) -0.00(0), shape factor 1.050)
  Q7: 1.09 - connects strings (connect value 10, shape factor 0.219)

Dan


(;GM[1]FF[3]
RU[Chinese]SZ[19]HA[0]KM[5.5]
PW[White]
PB[Black]
PL[White]
GN[White (W) vs. Black (B)]
DT[2006-01-27]
SY[Cgoban 1.9.12]TM[30:00(5x1:00)];
AW[ol][pl][rl][om][rm][on][qn][po][ro][pp][pq]
AB[cc][qc][dd][gd][jd][md][pd][dg][pg][dj][jj][nk][ok][pk][rk][nl][ql][dm][nm][qm][nn][pn][rn][no][oo][dp][gp][jp][mp][cq][nq]
)

[gunnar]

This is kind of similar to the issue solved in  http://lists.gnu.org/archive/html/gnugo-devel/2004-11/msg00062.html

[gunnar]

gunnar_7_9.1. Preparations.

[gunnar]

gunnar_7_9.2. Avoid playing into snapback.

[gunnar]

gunnar_7_9.2 adds a check that we don't consider moves playing into snapback in do_find_more_owl_attack_and_defense_moves(), just like we already do for send-two-return-one.

For this to work, however, it's necessary to apply the bugfix patch gunnar_7_9.1. The problem is that do_find_more_owl_attack_and_defense_moves() was called at both stackp==0 and stackp==1. This wasn't detected before because send_two_return_one() happened to just be silently disabled at stackp==1 (because the considered move had already been trymoved), while the new playing_into_snapback() fails an assertion instead. After gunnar_7_9.1 all these calls are done at stackp==0.

After these patches the previous test gives

  Q7:   2.33 (followup) - threatens to capture Q6
  Q7:   5.25 (followup) - threatens to capture R8
  Q7: not a safe move
[...]
Top moves:
1. T6  28.90
2. R5  25.80
3. P2  9.24
4. R4  6.93

I.e. completely successful.

[gunnar]

Breakage for gunnar_7_9.1:

nicklas2:102    PASS D4 [D4]
trevorc:1580    PASS C8 [C8]
nngs3:490       FAIL T13 [P15]
2 PASS
1 FAIL
Total nodes: 1678363021 3281413 12459622 (+0.023% +0.044% +0.0087%)

Additional breakage for gunnar_7_9.2:

Total nodes: 1678364837 3281419 12459623 (+0.00011% +0.00018% +8e-06%)

[gunnar]

I can't really say that I understand the details of the breakage for gunnar_7_9.1 but considering that this really is a bugfix and nothing terrible happened, I won't dig deeper.