Hi Aja,
I think (1) is what you are doing, if I understand correctly. Did you ever
try (2)? I imagine you might already have tried it (maybe it's bad?)
I tried only using CNN's probability.
Mix MM gamma with CNN is clery stronger. But using only CNN is still
stronger than Aya without CNN. I had thought too small CNN's
probability was bad for Mix, but it is not.
Aya with CNN vs Aya without CNN, 10000 playouts/move
winrate wins/games min p
0.956 284/297 (p < 0.0001 ) p*=50 Mix MM gamma with CNN's p
0.914 85/ 93 (p < 0.001 ) p*=50 Mix MM gamma with CNN's p
0.923 275/298 (p < 0.01 ) p*=50 Mix MM gamma with CNN's p
0.570 158/277 (p < 0.00001) p*=1 Use only CNN's p
0.720 206/286 (p < 0.00001) p*=10 Use only CNN's p
0.777 226/286 (p < 0.00001) p*=100 Use only CNN's p
p : CNN's probability p
p = result_CNN(pos(x,y));
if ( p < 0.0001 ) p = 0.0001;
p *= 50;
MM_gamma *= p
Regards,
Hiroshi Yamashita
----- Original Message -----
From: "Aja Huang" <[email protected]>
To: <[email protected]>
Sent: Monday, October 05, 2015 12:32 AM
Subject: Re: [Computer-go] Detlef's DCNN data
Hi Hiroshi,
On Fri, Sep 18, 2015 at 7:08 PM, Hiroshi Yamashita <[email protected]> wrote:
DCNN returns each move probabilty. I multiply it by 1000, and multiply
it by each move's rating. (r *= 1000 means multiply by 1000).
Aya calls DCNN when node is created. Aya makes 900 nodes in 10000
playouts. GTS 450 needs 17.4ms for a position. 900*17.4 = 15.6 sec
is needed. Aya needs 5 sec for 10000 playout without DCNN, and
20.6 sec with DCNN. So 4 times slower.
Did you try only using CNN's prior as described in our CNN paper? That is
to say,
For every legal move on the board,
1. Mix MM gamma values with CNN prior by treating CNN's probability p as
gamma value 1000 * p.
2. Use CNN's p.
I think (1) is what you are doing, if I understand correctly. Did you ever
try (2)? I imagine you might already have tried it (maybe it's bad?)
because it's a really easy experiment.
Aja
--------------------------------------------------------------------------------
_______________________________________________
Computer-go mailing list
[email protected]
http://computer-go.org/mailman/listinfo/computer-go
_______________________________________________
Computer-go mailing list
[email protected]
http://computer-go.org/mailman/listinfo/computer-go
