Hi,
if anyone still get email from this list:
DeepMind did it again
https://arxiv.org/abs/1911.08265
Best wishes
Magnus Persson
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
I think you misunderstood what I wrote,
if perfect play on 9x9 is 6000 Elo, then if the value function is 3000
Elo and MC eval is 2000 Elo with 1 second thinking time then it might
be that the combination of a value function and MC eval ends up being
2700 Elo. It could also be that it ends up
My guess is that there is some kind of threshold depending on the
relative strength of MC eval and the value function of the NN.
If the value function is stronger than MC eval I would guess MCEval
turns into a bad noisy feature with little benefit.
Depending on how strong MC eval is this thre
Remi twittered more details here (see the discussion with gghideki:
https://twitter.com/Remi_Coulom/status/969936332205318144
On 2018-03-05 10:16, Gian-Carlo Pascutto wrote:
On 28-02-18 07:13, Rémi Coulom wrote:
Hi,
I have just connected the newest version of Crazy Stone to CGOS. It
is base
Hi,
somebody asked me to post my views on 9x9 go on the list based on my
experience with correspondence go on OGS and little Golem.
I have been playing online correspondence tournaments since 2011 with
Valkyria which is a MCTS heavy playout MCTS using AMAF heavily tuned for
9x9. Also with
https://alphagoteach.deepmind.com/
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
https://arxiv.org/abs/1712.01815
Best
Magnus
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
The new Iteration 4 network has 25000 games added to the game pool based
on the first Iteration 3 network, but note there is a delay in learning
since the games are generated and added in the background.
From visual inspection it looks like the prior distribution looks more
finished. Really ba
Hi all time for the first update with real test from CGOS 9x9 using a
network trained on Odin AGZ exploratory self play games without network.
The I3 network is a network trained on 75000 games from I1 and I2, in
parallel as the games were played.
I1 is 5 very fast self play games where Odi
I think might be extra interesting since I train networks for 5.5 and
7.0 komi on 9x9, I have a theory that the 7.0 komi network can become
much stronger because
most jigo games has to played perfectly but with a move ordering the
puts maximum pressure on the opponent. With 5.5 komi most games e
Odin has now generated 5 games with 100 simulation and 90% prior for
a random move, and 1 games with 5000 simulations and 15% prior for a
random move, with temperature 1.
The network has been trained all the time and the loss function is now
dropping about 0.02 units every 10 hours or
Since the random games in my first iteration still only play legal moves
the network will learn to play uniformly legal moves, and will assign 0
probability for play moves on stones.
This happens after some hours of training for 9x9 and two days of
training for 19x19 in my case. Basically the
Hi, I also have some Zero stuff brewing since almost two days. Although
I depend on heavy playout MC-evaluation for self play.
I am using my Odin MC-engine as a base as it is. It can use a small AG
style policy network running on CPU implemented with Eigen (C++). It
does not have any value ne
Another way of seeing it: the evaluation of a program that can play
perfect would tend to give score close to 100% or 0 %.
As far as I know Alphago would stay close to 50% for a long time just
like most other programs (but with much higher correlation with the
actual outcome of course).
Magn
Hi, my 2 cent:
I think it is more or less redundant for the border. Alphago has a plane
for black, white and empty. So a border point is definitely different
anyway since it will have no features set in any plane. But all points
on the board with has a 1 set in one of the three b/w/e-planes.
http://www.yss-aya.com/cgos/19x19/cross/JuEYi.html
Best
Magnus
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
(3) CNN cannot learn exclusive-or function due to the ReLU
activation function, instead of traditional sigmoid (tangent
hyperbolic). CNN is good at approximating continuous (analog)
functions but Boolean (digital) ones.
Are you sure about that? I can imagine using two ReLU units to
construct
Yes, I actually thought "Wow, that will make Ingo happy!" when i read
that. :-)
-Magnus
On 2017-04-10 16:57, Ingo Althöfer wrote:
From my point of view the following part in the program
will be very interesting:
“Pair Go” — A game where one Chinese pro will play
against another...except they
https://deepmind.com/blog/exploring-mysteries-alphago/
-Magnus Persson
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
Is it a rumor or will ke Jie play Alpha-Go in April?
http://www.newsgd.com/news/2017-02/04/content_164638895.htm
Magnus Persson
___
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
Could be, but the next UEC Cup will be last... so maybe this will be
some kind of invitational event.
Magnus
On 2016-11-29 15:52, Michael Markefka wrote:
Just a wild guess, but I assume they'll go for the latest winner of
the UEC Cup as far as AI entrants are concerned.
On Tue, Nov 29, 2016 a
Yes, I think the important thing of the value function is to detect
moves that are very bad so that MC-eval does not have to sample more
than once for many variations.
If the evaluation function was trained on pro moves only, it would not
know what a bad move looks like. At least the evaluatio
Just one additional thing. If you plan to play online (CGOS) you need
have a buffer for lag.
That is if the opponent plays on until the board is filled, you might
end up playing another 100 moves. Even if your program responds locally
in 0.0001 seconds the lag of systems out of your control mi
Hi,
I just think I found a bug in twogtp 1.4.8 (windows), using the
-alternate flag and two programs that always resign if losing.
Basically the winner is wrong in the SGF files, and also in the logs.
The bug seems to happen only for odd games. It is easy to see because
when a program resigns
Hi all,
The current "Alan Turing Nines Title Tournament 2016" is also underway.
Note that it has 3 Rounds so the links shows the results from the 1st
round sofar.
Valkyria has won the last 3 9x9 Tournaments and I expect it to continue
to do well unless some of the really stron
Very interesting games!
My feeling is that darkforest rely too much on playing good shape,
and followed Kobayashis light moves too passively on the board.
Maybe it needs a better monte carlo evaluation (does it even have any?).
Zen on the other hand took every opportunity to attack white wildly
itself.
This is also my experience from playing Correspondence go with Valkyria
on 9x9.
The program is very far from perfect play, but with long thinking times
one
must play patient but sharp. Setting traps does not work, but if one is
luckily the program
will go into some position which it
Hi!
I think the technique of hashing move pairs from the search tree and
reuse them in
the playouts if the context matches, could plausibly be the major
improvement of Alphago that
we witnessed today.
Another thing I noticed is that Alphago does not use any statistics from
the playouts other
Congratulations!
What I find most impressive is the engineering effort, combining so many
different parts, which even standalone would be a strong program.
I think the design philosophy of using 3 different sources of "go
playing" strength is great in it self (and if you read the paper there
playout made it work well enough to compete for
some time even with the first UCT programs.
Magnus Persson (Viking/Valkyria)
On 2016-01-11 17:15, Hendrik Baier wrote:
Hi Xavier,
I have considered this, but even level-2 Nested MCTS requires quite a
bit of thinking time. Let's say you c
I once tried to use 32bit hashes only. It might work for super ko
detection, but if you use it for transposition detection storing a lot
of poisitions you will get collisions now and then. So my reasoning is:
if 32 bits almost works then using another 32 bits will work.
Best
Magnus
On 2015-0
Interesting position.
I am not sure how I exactly handle this is in Valkyria. My code for
allowing or disallowing suicide is very complex.
As long as "special case"-handling is only executed when necessary.
In this case the pattern with two stones forming a possible eye would be
At least the 13x13 server is working now.
-Magnus
On 2015-03-04 09:18, Urban Hafner wrote:
On Tue, Mar 3, 2015 at 9:57 PM, Joshua Shriver
wrote:
Thanks for the heads up *sigh*
I really am contemplating re-writing CGOS from scratch. TCL is
just
quirky as hell. Plus would make updates easier
I forgot to turn of automatic Power off in Windows so after an hour my
computer hibernated. I had started Valkyria again this morning (now
using 6 threads) and then CGOS seemed to recover.
Maybe CGOS froze because of this?
Magnus
___
Computer-go
I just connected Valkyria on 13x13!
:)
On 2015-01-15 20:20, Urban Hafner wrote:
On Thu, Jan 15, 2015 at 6:45 PM, Christoph Birk
wrote:
On Jan 15, 2015, at 9:35 AM, Joshua Shriver
wrote:
Aye I'm still tinkering with it, and trying to get anchors on.
Still
having issues. :(
The
Hi all,
I am working on a successor to Valkyria written in C++.
I agree with Petr here. I am still tweaking the basic engine and I have
still lot to fix. With a sloppy implementation initially I have 20 kpps
for 9x9 and 4.5 kpps for 19x19.
When the basic stuff is running I will build feature
36 matches
Mail list logo
