Quoting Don Dailey <[email protected]>:
Are we talking about opening books or learning systems?
If you have a learning system on CGOS that builds a book as it plays, that
is a completely different animal.
Years ago in chess it was shown how to crush any opponent in a long enough
match (even with a weaker program) by using an opening book learning system.
I'm simplifying this a bit, but the general concept is to play the
openings that you are winning with and to NOT play the ones you are losing
with. If you play enough games you can automate the process of figuring
out which sequence of moves to play in order to beat a specific opponent or
a set of opponents.
My first programs on CGOS did that. Against a static opponent it is
easy to improve a book automatically or manually and beat it.
The problem is that this book against other programs will not
generalize well because todays program are playing far from perfect.
Partially because it learns moves that provokes mistakes in the
opponent. Partially because it avoids correct lines that it cannot
handle.
CGOS is the best source for progress in openings. But not playing
simple opponents. If I can figure out how to deal with bonobot it will
not help against Zen.
The hardest problem is to get enough solid games to work with. Going a
ply deeper in the book means you need to have 10 times more games or
so to have a chance to see what lines are good and not. And the better
the book gets the harder it is to truly understand the games. Good 9x9
go games have a higher branching factor later in the game because you
get into fight where you need to play forcing moves and ko threats in
an exact order to get it right.
Here is some more food for though. This is the distribution of the
number of positions for each depth and the
"openingbook-branching-factor".
Depth/Num positions/OB-branching-factor
0 1
1 6 6
2 13 2.2
3 99 7.6
4 144 1.5
5 342 2.4
6 416 1.2
7 604 1.5
8 590 1.0
9 744 1.3
10 629 0.8
11 692 1.1
12 555 0.8
13 556 1.0
14 490 0.9
16 404 0.8
17 379 0.9
18 254 0.7
19 250 1.0
20 203 0.8
21 195 1.0
22 119 0.6
23 117 1.0
24 76 0.6
25 89 1.2
26 51 0.6
27 50 1.0
28 33 0.7
29 28 0.8
30 18 0.6
31 20 1.1
32 13 0.7
33 10 0.8
34 9 0.9
35 6 0.7
36 4 0.7
37 4 1.0
38 4 1.0
I stopped counting positions after depth 38. There is a single game
reaching 70 ply.
The highest number of positions are for depth 9. Upto this depth there
are 2959 positions, but in total there are 8246 position so the
majority of positions are very deep. This is because to consistently
outplay single strong opponents the lines much be very long. But these
lines will not be repeated against new strong opponents with other
biases in their search.
This is not evidence of anything but my feeling is that a book
claiming it has solved 9x9 go has to be very deep (unless engines get
so strong in search they do not need search) and the branching factor
much be much higher than 1.0 to be useful in practice.
Best
Magnus
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
