Hi,
If I understood correctly you would try to use a program using value net
with (let's say 2000 playouts) in selfplay? Using only one result, or
Yes. 2000 playouts/move MCTS with policy net and value net.
doing some games per position? Or are you thinking of using only the win
I thought one game per position, but some games per position looks nice option.
doing some games per position? Or are you thinking of using only the win
percentage such a program gives from his own mixing of SL network,
In my experience, game result is better than win percentage.
usage per game (at least if Rn3.3-4c is Ray on CGOS with 4 cores, NG04b
Oh NG04b is oakfoam. AlphaGo RL is about 2800 (CGOS BayesElo).
So around this rating program seems nice. And many computers don't have GPU.
To calculate DCNN on CPU, maybe we can not use big network(filter 192), but
smaller one(filter 64 or 32).
Hiroshi Yamashita
----- Original Message -----
From: "Detlef Schmicker" <d...@physik.de>
To: <computer-go@computer-go.org>
Sent: Friday, January 06, 2017 7:20 PM
Subject: Re: [Computer-go] it's alphago (How to get a strong value network)
Hi,
this sounds interesting! AlphaGo paper plays only with RL network, if I
understood correctly. If we start this huge approach we should try to
carefully discuss the way (and hopefully get some hints from people
tried with much computational power :)
If I understood correctly you would try to use a program using value net
with (let's say 2000 playouts) in selfplay? Using only one result, or
doing some games per position? Or are you thinking of using only the win
percentage such a program gives from his own mixing of SL network,
search and value net?
By the way to make some promotion :) oakfoam is not far away from Ray
for this kind of approach, where you will probably try to reduce cpu/gpu
usage per game (at least if Rn3.3-4c is Ray on CGOS with 4 cores, NG04b
is oakfoam on CGOS with 10k and saving GPU usage by using only 50% of GX970)
Detlef
Am 06.01.2017 um 10:39 schrieb Hiroshi Yamashita:
If value net is the most important part for over pro level, the problem
is making strong selfplay games.
1. make 30 million selfplay games.
2. make value net.
3. use this value net for selfplay program.
4. go to (1)
I don't know when the progress will stop by this loop.
But if once strong enough selfplay games are published, everyone can
make pro level program.
30 million is big number. It needs many computers.
Computer Go community may be able to share this work.
I can offer Aya, it is not open-source though. Maybe Ray(strongest open
source so far) is better choice.
Thanks,
Hiroshi Yamashita
----- Original Message ----- From: <fotl...@smart-games.com>
To: <computer-go@computer-go.org>
Sent: Friday, January 06, 2017 4:50 PM
Subject: Re: [Computer-go] it's alphago
Competitive with Alpha-go, one developer, not possible. I do think it is
possible to make a pro level program with one person or a small team.
Look at Deep Zen and Aya for example. I expect I’ll get there (pro
level) with Many Faces as well.
David
_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
