On Jun 29, 2011, at 2:39 PM, Brian Sheppard wrote:
>Such a table has 47 million entries in 19x19 go and over half a
million in 9x9 go. It require an enormous amount of data to fill
such a table with enough samples to not be almost meaningless
statistically.
That's true, but... But there are two other aspects to the math.
1) Classifiers are expensive to train
Note that the paper compares on fixed trials rather than fixed time.
The first experiment is on fixed trials; the rest are on fixed time.
A network with N outputs, H hidden nodes and K inputs has H * (N +
K) weights. Here, K is the board size (9x9 = 81), and N varied from
10 to 40, IIRC. Orego is training those weights on every position in
the playout (That is, roughly K moves worst case). Each update
involves dozens of floating point calculation.
We found that hidden units hurt, even with fixed playouts. As the
paper explains, a linear classifier performs even better (on a time
basis) than a sigmoid network with no hidden units. The number of
weights trained on each playout move is only 3, each of which only
requires a few flops.
Basically: RAVE uses (a lot) less computation and gets a result that
is better. (But maybe not much better; hard to tell because there is
no comparison to Orego+RAVE.)
Yes, RAVE is much better. :-)
Peter Drake
http://www.lclark.edu/~drake/
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
