[Computer-go] Computer Analysis of Master Games ?!

Petr Baudis pasky at ucw.cz
Tue Aug 10 05:00:04 PDT 2010


On Tue, Aug 10, 2010 at 11:34:01AM +0200, "Ingo Althöfer" wrote:
> 
> > My proposal would be to include also a score estimator curve.
> > Both diagrams together (win rates and scores) give a better clue
> > about what is going on.
> 
> I have taken another small sample:
> 4 recent games between Andy Liu (alias BigBadWolf on kGS) and
> Curtis Tang (alias cjheaven on KGS)
> 
> and generated both win rate and score chart for the games.
> 
> http://www.althofer.de/computer-analysis-cjh-bbw.html
> 
> For game 2 you can see how Monte-Carlo went astray in the (very) end.
> (Black won that game.)
> 
> Only for game 4 both diagrams together give a trustful picture.

Pachi's perspective. Diagrams are attached, all are
eventual-winner-to-play.

Game 1:

Pachi did not appreciate B6, like some observers - at that point,
winrate fell sharply. It recovered the moment black ignored ko and
played S18. From that moment on, it always stayed >50% for black.
Like with manyfaces, there is a sharp drop after move 185 - Pachi did
not appreciate black letting M17 live, thinking of that as being
unneccessarily risky, even though white can never cut off the black
group of course - but Pachi sees that and the winrate stays >50%.
Another smaller drop is caused by G18, I believe because in high number
of simulations Pachi does not connect all of its black stones in time
to avoid connect-and-die.

Game 2:

This is a very interesting game! Pachi agrees that this was a very even
game all throughout, but its winrate stays near 50% until the very end.
The final position is half a point win for black, but why Pachi does not
see it? Because MCTS is actually really poor in endgame in close games!
This is equivalent to discerning semeai sequences - all endgame moves
need to be played in precise order, and this includes dame. Black must
fill all dame before playing into its own territory, and that's actually
very hard to do and requires extremely deep tree branches! If black
would be 10 points ahead, straying and playing some moves in its
territory does not matter - if the game is even, it's as likely that
black will play less dame as the opposite.

Game 3:

Pachi did not understand this game at all. The graph is similar to
ManyFaces, but without the final improvement for white. First confusion
came around B12 - Pachi needed about 40k simulations to find that move,
and still had trouble realizing that if black tries to kill white, his
corner will die. After E13, it got happier. B7 was the cause for the
second winrate fall - Pachi preferred the flimsy notion that the black
group might just die off completely later in the simulations. Overally,
white's winrate does not go up since Pachi of course does not realize
bottom right corner is dead - it lives in around 50% of games since
Pachi likes to play off W nakade stones and create living shape for B.
After manually settling the corner, the winrate for w is >60%.

Game 4:

The winrate graph seems to correspond well to the game. Pressed for a
comment, Pachi thinks J18 is too slow and black shold C2 instead of L17.

-- 
				Petr "Pasky" Baudis
The true meaning of life is to plant a tree under whose shade
you will never sit.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cjheaven-bigbadwolf-1-B5k.png
Type: image/png
Size: 5390 bytes
Desc: not available
URL: <http://computer-go.org/pipermail/computer-go/attachments/20100810/d0bca1fc/attachment.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cjheaven-bigbadwolf-2-B5k.png
Type: image/png
Size: 4804 bytes
Desc: not available
URL: <http://computer-go.org/pipermail/computer-go/attachments/20100810/d0bca1fc/attachment-0001.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cjheaven-bigbadwolf-3-W5k.png
Type: image/png
Size: 5111 bytes
Desc: not available
URL: <http://computer-go.org/pipermail/computer-go/attachments/20100810/d0bca1fc/attachment-0002.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cjheaven-bigbadwolf-4-W5k.png
Type: image/png
Size: 5312 bytes
Desc: not available
URL: <http://computer-go.org/pipermail/computer-go/attachments/20100810/d0bca1fc/attachment-0003.png>


More information about the Computer-go mailing list