Now when Remi said 7d this rank is based on internet servers like KGS or EGF ranking system?
Crazystone
-
Rémi
- Lives with ko
- Posts: 170
- Joined: Sat Jan 14, 2012 4:11 pm
- Rank: KGS 4 kyu
- GD Posts: 0
- Has thanked: 32 times
- Been thanked: 119 times
- Contact:
Re: Crazystone
Krama wrote:I can't wait to have a 7d go buddy at my disposal
Now when Remi said 7d this rank is based on internet servers like KGS or EGF ranking system?
CrazySton7 reached 7d on KGS, playing on a 4-core machine at 2 seconds per move:
https://www.gokgs.com/gameArchives.jsp? ... 16&month=4
- ez4u
- Oza
- Posts: 2414
- Joined: Wed Feb 23, 2011 10:15 pm
- Rank: Jp 6 dan
- GD Posts: 0
- KGS: ez4u
- Location: Tokyo, Japan
- Has thanked: 2351 times
- Been thanked: 1332 times
Re: Crazystone
Rémi wrote:Krama wrote:I can't wait to have a 7d go buddy at my disposal
Now when Remi said 7d this rank is based on internet servers like KGS or EGF ranking system?
CrazySton7 reached 7d on KGS, playing on a 4-core machine at 2 seconds per move:
https://www.gokgs.com/gameArchives.jsp? ... 16&month=4
Remi, Why did you play it so quickly? On that level of hardware, what difference would 10 seconds per move or 20 seconds per move have made?
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21
-
Rémi
- Lives with ko
- Posts: 170
- Joined: Sat Jan 14, 2012 4:11 pm
- Rank: KGS 4 kyu
- GD Posts: 0
- Has thanked: 32 times
- Been thanked: 119 times
- Contact:
Re: Crazystone
ez4u wrote:Remi, Why did you play it so quickly? On that level of hardware, what difference would 10 seconds per move or 20 seconds per move have made?
I was calibrating weak levels for the commercial release. I did not expect it would reach 7d. Next week I'll connect Crazy Stone on a powerful server. I believe there is a chance that it may reach 8d. We'll see.
-
pookpooi
- Lives in sente
- Posts: 727
- Joined: Sat Aug 21, 2010 12:26 pm
- GD Posts: 10
- Has thanked: 44 times
- Been thanked: 218 times
Re: Crazystone
Rémi wrote:ez4u wrote:Remi, Why did you play it so quickly? On that level of hardware, what difference would 10 seconds per move or 20 seconds per move have made?
I was calibrating weak levels for the commercial release. I did not expect it would reach 7d. Next week I'll connect Crazy Stone on a powerful server. I believe there is a chance that it may reach 8d. We'll see.
On KGS? I'll stay awake for the whole week!!!
-
Satorian
- Dies with sente
- Posts: 93
- Joined: Sun Jan 22, 2012 7:11 am
- Rank: Total beginner
- GD Posts: 0
- Universal go server handle: Satorian
- Has thanked: 26 times
- Been thanked: 13 times
Re: Crazystone
I've got a desktop running an i7 4790k at 4 GHz, and a laptop running an i7 6700HQ. Looking forward to how well they do. 
The desktop is also equipped with a GTX 980Ti, while the laptop has got a 960M, both with current CUDA installations. Wouldn't mind seeing a GPU-implementation of CS' DCNN at some point.
The desktop is also equipped with a GTX 980Ti, while the laptop has got a 960M, both with current CUDA installations. Wouldn't mind seeing a GPU-implementation of CS' DCNN at some point.
Re: Crazystone
Three quick questions @Remi:
1. Do you get significant improvement from your neural networks in 13x13 and 9x9? If so, is it correct to assume that the training of the small board variants was completely separate from the training of the standard size net?
2. Will the play of the weak pure DCNN levels adjust to komi?
3. Does CSDL have an analogue to the AlphaGo value network or just to the policy network at the moment?
1. Do you get significant improvement from your neural networks in 13x13 and 9x9? If so, is it correct to assume that the training of the small board variants was completely separate from the training of the standard size net?
2. Will the play of the weak pure DCNN levels adjust to komi?
3. Does CSDL have an analogue to the AlphaGo value network or just to the policy network at the moment?
-
Pippen
- Lives in gote
- Posts: 677
- Joined: Thu Sep 16, 2010 3:34 pm
- GD Posts: 0
- KGS: 2d
- Has thanked: 6 times
- Been thanked: 31 times
Re: Crazystone
I will wait and see how strong CS will be on weaker machines. I have a laptop with duo core, 2,4 GHz and 4GB RAM and I am 2d-3d strong, so I have some doubt if CS can be at least one full stone stronger than me (I want to use it for analysis and as a teacher).
-
Satorian
- Dies with sente
- Posts: 93
- Joined: Sun Jan 22, 2012 7:11 am
- Rank: Total beginner
- GD Posts: 0
- Universal go server handle: Satorian
- Has thanked: 26 times
- Been thanked: 13 times
Re: Crazystone
Pippen wrote:I will wait and see how strong CS will be on weaker machines. I have a laptop with duo core, 2,4 GHz and 4GB RAM and I am 2d-3d strong, so I have some doubt if CS can be at least one full stone stronger than me (I want to use it for analysis and as a teacher).
You can set fixed time limits per move or per game IIRC for whole game analysis. So, if you don't need that 7-8d strength live while playing, your configuration should be more than enough to be able to analyze games at that level when you give CS 2-3h of time.
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: Crazystone
Here comes Rémi's kind answer to a similar question of mine, related to "Analyse Current Game" = analyse the current position = show the next move's candidates.
This feature (for a single move) is not the same as "Analyse Game Record ..." (for every move contained in a SGF's main line), which is mentioned in the posting above, but I think that Rémi's core message will apply to that as well.
"The analysis will be of the same quality for the same number of playouts, whatever the hardware. Of course, the same number of playouts will be faster on a more powerful machine."
So you will be able to get a high-level analysis done be CS, if time is no limiting factor for you.
This feature (for a single move) is not the same as "Analyse Game Record ..." (for every move contained in a SGF's main line), which is mentioned in the posting above, but I think that Rémi's core message will apply to that as well.
"The analysis will be of the same quality for the same number of playouts, whatever the hardware. Of course, the same number of playouts will be faster on a more powerful machine."
So you will be able to get a high-level analysis done be CS, if time is no limiting factor for you.
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)
-
Pippen
- Lives in gote
- Posts: 677
- Joined: Thu Sep 16, 2010 3:34 pm
- GD Posts: 0
- KGS: 2d
- Has thanked: 6 times
- Been thanked: 31 times
Re: Crazystone
Cassandra wrote:"The analysis will be of the same quality for the same number of playouts, whatever the hardware. Of course, the same number of playouts will be faster on a more powerful machine."
So you will be able to get a high-level analysis done be CS, if time is no limiting factor for you.
Thx for the info. If Remi can confirm that it'd be a huge argument to buy CS.
-
pragmaticleas
- Dies in gote
- Posts: 51
- Joined: Sun Sep 16, 2012 4:18 pm
- GD Posts: 0
- Has thanked: 3 times
- Been thanked: 5 times
Re: Crazystone
Pippen wrote:Cassandra wrote:"The analysis will be of the same quality for the same number of playouts, whatever the hardware. Of course, the same number of playouts will be faster on a more powerful machine."
So you will be able to get a high-level analysis done be CS, if time is no limiting factor for you.
Thx for the info. If Remi can confirm that it'd be a huge argument to buy CS.
You can check out this thread where Remi seemed to have confirmed it, but you may want another confirmation since this is another Crazy Stone version, and I am not sure if there is a timed mode for the new version:
http://www.lifein19x19.com/forum/viewto ... 1&start=40
Rémi wrote:In terms of speed, Crazy stone does about 250 playouts / second / core on the transformer prime, so 1000 playouts / second on 4 cores. This compares to about 2000 playouts / second /core on a PC. Except for speed, the program is otherwise identical.
Rémi
- Cassandra
- Lives in sente
- Posts: 1326
- Joined: Wed Apr 28, 2010 11:33 am
- Rank: German 1 Kyu
- GD Posts: 0
- Has thanked: 14 times
- Been thanked: 153 times
Re: Crazystone
pragmaticleas wrote:You can check out this thread where Remi seemed to have confirmed it, but you may want another confirmation since this is another Crazy Stone version, and I am not sure if there is a timed mode for the new version:
CS 2012 has it, CS 2013 has it. CS 2016 screenshots look like the same.
Analysis features are explicitely mentioned in the Unbalance announcement.
Why should they make a step back ?
In my opinion, you will use the "Analyse Game Record ..." feature to identify the "hot spots" of a game seen by CS (e.g. where CS's favourite move has a much better winning percentage than the really played one).
CS 2012 has about 30.000 playouts per move with the setting "depth = high". (Lasted about 20 to 30 minutes in total on my low-level laptop).
I did not try the "Total time" option myself, but the maximum time that can be chosen is 8 hours !
"Analyse Current Game" would be used to get an idea of the moves CS has thought about (additionally maybe the histogram) in a special position, and to try out variations.
This analysis for one special move will last until you stop it (e.g. by going to the next move played).
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)
Igo Hatsuyōron #120 (really solved by KataGo)