[Computer-go] Congratulations to Zen!

valkyria at phmp.se valkyria at phmp.se
Tue Apr 5 05:57:49 PDT 2011


Quoting Jean-loup Gailly <jloup at gailly.net>:

> Magnus writes:
>
>> Further, I would imagine many jigos could be quite peaceful short games
>
> What makes you think that optimal play should be peaceful? I speculate
> that it should often be very complex, full of kos and tenukis, and
> very difficult to understand for humans ("I play here because it gains
> 1 point 20 moves later"). Some games might be peaceful, but a perfect
> player can select arbitrarily complex sequences as long as they still
> achieve the perfect result. There are a lot more complex games than
> peaceful games.

I think you are right. Maybe it is possible to create a book that  
leads to short jigos within the tree of "perfect games" but of course  
a strong book should prefer longer and complicated lines without  
making mistakes.

My speculation was that short lines for jigo might be missing in a  
book for 7.5, but I do not think that can explain that lack of jigos  
in the tournament.

I started a proper selfplay experiment with komi 7.0 with thinking  
time clearly longer than for CGOS to get high quality games.

So far for 38 games there were 7 Jigo (18%).
B won 11 28%
W won 20 52%

This seems to be the same patterns as in the tournament, where white  
has a clear advantage but less so because more jigo thanks to selfplay  
+ much longer thinking times.

Actually the games were played with following first moves forced  
systematically, as if a 1 ply uniform opening book

Move #B Won/#Games

E5 4/8
E4 1/6
F4 3/6
E3 3/6 Jigo 1
F3 0/6 Jigo 5
G3 0/6 jigo 1

Moves near the center are stronger as expected, but all jigos occurred  
with starting moves close to the edge.

I know nothing can be concluded from this data set. But it is fun to  
speculate.
Maybe E5 is a losing move for black! But it is such a complicated move  
and leads to fighting that Jigo are rare and a move such is F3 gives  
Jigo with perfect play.

Speaking against this may be that with komi 5.5 the moves near the  
edge does not seems so strong for black. But so far I cannot rule out  
that all those opening moves are a win for black with komi 5.5. Anyway  
the opening in 9x9 remains mysterious in most parts, no matter how  
much data I collect....

Or is that with slow starting moves blacks only hope is jigo and thus  
achieves it more often?


I used to be skeptical against integral komi but now I feel that it  
makes blacks play more interesting than it is with komi 7.5.

I wonder if Don would be willing to start a CGOS server with 9x9 komi  
7.0? And would people then be willing to play on it? I think in the  
long term this is the best way getting an idea how Jigo impacts on  
play. It might even help books development for 7.5 because black lines  
that loses with no chance of jigo might then be rejected, and the  
remaining lines is perhaps also strong for 7.5. (Speculation).

I also checked game length for the jigos picking the last move that  
took a real point:

70 53 60 71 77 52 73
Mean 65

All other games were resigned on 20% win rate (in prefer many games in  
self play, compared to be sure that the right player wins), and were  
on average much shorter.

Resigned games
75 35 38 53 58 45 46 51 22 44 38 44 54

Mean 46

So it seems to be no evidence for any "quick" Jigos for this  
preliminary result.

Magnus






More information about the Computer-go mailing list