Thanks everyone for the answers regarding playout terminations. I still
have my suspicions regarding how artificial game length bounds affect
the position evaluation (I'd expect the values to fluctuate with length,
so arbitrary bounds would result in arbitrary differences).
For the moment, however, I'd like to focus on the variations wrt
pseudo-eyes (positions not to be filled during playouts). All of these
are known to be wrong in some cases, the main argument in favour of
pseudo-eyes being efficiency.
Since ruling out fill-ins introduces not only biases, but actual blind spots
into the playouts, something even heavy playouts otherwise try to avoid
(they prioritize "good" moves instead of eliminating "bad" moves) it is
important that the definition of pseudo-eyes errs on the safe side ("safe"
here meaning random play - any move is possible): they may permit
erroneous fill-ins, but must not rule out necessary fill-ins (only definitely
bad fill-ins may be ruled out).
Within that constraint, the goal is to rule out as many bad fill-ins as
possible without ruling out any good fill-ins. And all three variations
listed earlier (Gobble, Olga, Oleg)
http://computer-go.org/pipermail/computer-go/2008-October/016549.html
have their issues with that. For instance, consider the following position:
(
;
FF[1]
GM[1]
SZ[19]
AP[Jago:Version 5.0]
AB[ji][kj][jk][ij][jl][hj][jh][lj][jg][kf][ke][le][me][ne][of][og][oh][oi][ni][mj][nf][kg][nj][jm][jn][jo][jp][ip][hp][gp][fp][fo][fn][fm][fl][fk][fj][gj][gi][gh][hg][ig][ok][gf][pl][fe]
AW[ki][kh][li][lg][mh][lf][mf][ng][nh][kk][lk][kl][mi][mk][nk][ik][hk][gk][gl][gm][gn][il][im][in][io][go][hm][ho][jd][kd][ld][md][nd][od][pe][pf][ph][pg][pi][pj][ei][ej][ek][el][em][en][eo][ep][eq][fq][gq][hq][iq][jq][kq][ko][kp][kn][km][oj][ii][fi][fh][fg][gg][hf][if][jf][je][ih][oe]
C[black to move.
Gobble: 'a' is not an eye => fill it and die
Olga: 'a' is an eye, but would cease to be if 'b' was white
Oleg: 'a' is not an eye, but would be if 'b' was black]
L[jj][hh][nk][oj]
GN[playout-eyes]
)
If I set this up correctly, the black center group is "unconditionally" alive,
even if it was white's turn, while playouts with Gobble-style pseudo-eyes
will rate it as unconditionally dead (and hence the white strings 'c' and 'd'
as unconditionally alive, independent of conditions on the outside). Playouts
with Olga-style pseudo-eyes would fare better, but would be fooled into
considering fill-in at 'a' after white 'b'. Playouts with Oleg-style pseudo-eyes
have a chance of finding the black group alive, in case of black 'b' first.
Did I set this up right? And why does no-one else seem worried about it?-)
Claus
PS In case the sgf-labelling isn't standard, 'a' is K10/center, 'b' is H12.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
