It is currently Sat Apr 20, 2019 11:14 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 47 posts ]  Go to page 1, 2, 3  Next
Author Message
Offline
 Post subject: Elf OpenGo paper released
Post #1 Posted: Wed Feb 13, 2019 3:04 am 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
https://arxiv.org/abs/1902.04522

https://arxiv.org/pdf/1902.04522.pdf


This post by Uberdude was liked by 4 people: Bill Spight, Bonobo, Calvin Clark, Marcel Grünauer
Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #2 Posted: Wed Feb 13, 2019 3:32 am 
Lives in gote

Posts: 552
Location: Vienna, Austria
Liked others: 246
Was liked: 285
And ELFv2 is released. https://github.com/leela-zero/leela-zero/issues/2212

You can download it for Leela Zero at https://share.weiyun.com/5gLzPL5

This network is 20x256 (ELFv1 is 20x224).


Last edited by Marcel Grünauer on Wed Feb 13, 2019 3:54 am, edited 2 times in total.

This post by Marcel Grünauer was liked by: Bonobo
Top
 Profile  
 
 Post subject: Re: Elf OpenGo paper released
Post #3 Posted: Wed Feb 13, 2019 3:48 am 
had to happen... now what are the LZ crowd gonna do about this?!

it's all Geek to me, but i'd love to see C2's move that nearly knocked Elf off her perch - from the authors' remarks about ladders, it looks like Elf walked under one and nearly paid the price, but made a stunning recovery.
Attachment:
elfvc2.png
elfvc2.png [ 66.58 KiB | Viewed 1399 times ]

Top
  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #4 Posted: Wed Feb 13, 2019 4:01 am 
Lives with ko
User avatar

Posts: 293
Location: Deutschland
Liked others: 246
Was liked: 122
Rank: EGF 4 kyu
At this moment in time, it looks like ELF v.2 (05dbca15) is losing solidly against the strongest LZ network. The match is not over, of course:

Attachment:
Capture.JPG
Capture.JPG [ 36.77 KiB | Viewed 1386 times ]

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #5 Posted: Wed Feb 13, 2019 4:01 am 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
jaca wrote:
but i'd love to see C2's move that nearly knocked Elf off her perch - from the authors' remarks about ladders, it looks like Elf walked under one and nearly paid the price, but made a stunning recovery.


Presumably it was this 4th game, which we talked about before (black 0.47875 is very close to the near -0.5 of graph):


Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #6 Posted: Wed Feb 13, 2019 4:04 am 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Charlie wrote:
At this moment in time, it looks like ELF v.2 (05dbca15) is losing solidly against the strongest LZ network. The match is not over, of course:

In previous matches Elf has lost more of the early games because it tends to lose games to ladder blunders which are short games, but does relatively better at games which go to counting and they take longer.

Top
 Profile  
 
 Post subject: Re: Elf OpenGo paper released
Post #7 Posted: Wed Feb 13, 2019 4:13 am 
proof, as if any were needed, that 43 is the Meaning of Life :study:

a study group type question, if i may (perhaps best answered over there): that joseki in the bottom right - i vaguely remember St Michael talking about it onetime, but can't remember what he said or where he said it; it looks wierd to me, for what's the point in jumping with keima if white's only going to butt heads against it and black more or less has to play where she would have done anyway with the more (to me, at least) natural-looking press instead? i suppose there must be some advantage to the keima that doesn't become obvious until several hundred moves later...

Top
  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #8 Posted: Wed Feb 13, 2019 6:36 am 
Beginner

Posts: 5
Liked others: 1
Was liked: 0
What happened in these two games?

W: ELF V2
B: LeelaZero #204

http://zero.sjeng.org/viewmatch/c327835 ... viewer=wgo
http://zero.sjeng.org/viewmatch/7b181c4 ... viewer=wgo

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #9 Posted: Wed Feb 13, 2019 8:28 am 
Lives with ko

Posts: 205
Liked others: 41
Was liked: 62
KGS: lepore
jaca wrote:
...what's the point in jumping with keima if white's only going to butt heads against it and black more or less has to play where she would have done anyway with the more (to me, at least) natural-looking press instead? i suppose there must be some advantage to the keima that doesn't become obvious until several hundred moves later...


I hope I don't get in trouble for answering in the wrong place:

- I believe the keima increases black's ability to take sente. Which seems to be a thing now.
- I also don't think it is a given that white is going to butt heads. I've seen O-02 played as well, and the variations can get complicated.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #10 Posted: Wed Feb 13, 2019 8:51 am 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
tigerman wrote:


Are there bad parameters? (LZ test game files only seem to show params for one colour, or do we assume both colours have the same parameters expcet for network file?) Looks like those training games where first x moves have chance to be random. Also malicious/modified/badly-configured clients is a possibility, it's been a cause for concern on the LZ project for a while.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #11 Posted: Wed Feb 13, 2019 9:01 am 
Beginner

Posts: 5
Liked others: 1
Was liked: 0
I agree that it was probably due to client misconfigurations. Taking these two games out, ELF v2 vs. LZ #204 was a measly 34-60, or 36.17% games won. Not great, at least in this round of experiments.

All match games in the round can be viewed at http://zero.sjeng.org/match-games/5c63e ... 029e7a2390.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #12 Posted: Wed Feb 13, 2019 9:06 am 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Not so surprising give it is equal visits so the 40-block LZ will likely use more than twice the time as 20-block Elfv2. Here is a test match with the same equal visits setup of LZ 200 vs Elf v1, Elfv1 won 33.5% then.

https://github.com/leela-zero/leela-zero/issues/2212. A Chinese user @atoutw is doing an 120 sec a move test on 2080Ti Elfv2 vs LZ 204.

https://cloudygo.com/ringmaster/minigo- ... lf.report/ A minigo v15 (20 block) vs elf v2 test match on 3200 visits

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #13 Posted: Wed Feb 13, 2019 9:17 am 
Beginner

Posts: 5
Liked others: 1
Was liked: 0
Can anyone confirm that ELF can correctly read ladders now (perhaps with the right number of playouts or visits)? I saw ELF v2 still making questionable ladder moves in a couple of games against LZ #204.

I was so hoping to see the machine play a double ladder breaker some day. :-)

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #14 Posted: Wed Feb 13, 2019 9:41 am 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
LZ too still has troubles with ladders, though is better than Elf. Here's a ladder fail of LZ 198: forum/viewtopic.php?p=240305#p240305. I'll give it a try with LZ 204 and 50k+ playouts later.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #15 Posted: Wed Feb 13, 2019 9:49 am 
Beginner

Posts: 5
Liked others: 1
Was liked: 0
The latest-

ELF v2 ladder fail: http://zero.sjeng.org/viewmatch/49314be ... viewer=wgo

LZ #204 ladder fail: http://zero.sjeng.org/viewmatch/12f4fa2 ... viewer=wgo

Minigo v16-1144 ladder fail: http://zero.sjeng.org/viewmatch/d8696d1 ... viewer=wgo

Parameters are -p 0 -v 1600 -r 5 -m 0 -t 1 -d --noponder.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #16 Posted: Wed Feb 13, 2019 10:55 am 
Lives in sente

Posts: 1328
Liked others: 615
Was liked: 412
Rank: AGA 3k KGS 2k
GD Posts: 61
KGS: dfan
tigerman wrote:
Can anyone confirm that ELF can correctly read ladders now (perhaps with the right number of playouts or visits)? I saw ELF v2 still making questionable ladder moves in a couple of games against LZ #204.

The paper explicitly mentions that it still has trouble with ladders.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #17 Posted: Wed Feb 13, 2019 11:17 am 
Beginner

Posts: 5
Liked others: 1
Was liked: 0
Thanks dfan! I found the part that you mentioned in the paper:

Quote:
Ladders create a predictable sequence of moves for each player that can span the entire board. In contrast to humans, Silver et al. (2017) observes that deep RL models learn these tactics very late in training. To gain further insight into this phenomenon, we curate a dataset of 100 ladder scenarios (as described in Appendix B.2) and evaluate the model’s ladder handling abilities throughout training.

In general, ladder play improves with a fixed learning rate and degrades after the learning rate is reduced before once again recovering. While the network improves on ladder moves, it still makes mistakes even late in training. In general, increasing the number of MCTS rollouts from 1,600 to 6,400 improves ladder play, but mistakes are still made.

Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #18 Posted: Wed Feb 13, 2019 11:53 am 
Oza
User avatar

Posts: 2124
Location: Germany
Liked others: 7676
Was liked: 876
Rank: OGS SDK
Universal go server handle: trohde
Just to expand on the above info (without me understanding a word of all that AI blurb :lol: ) …

Qucheng Gong in the Facebook Group “Go (Baduk, Weiqi) Players on Facebook”:

Quote:
We have released ELF Opengo's all selfplay data and checkpoint models, a paper, a windows executable, as well as its analysis on all professional games ever played!
https://arxiv.org/abs/1902.04522
https://facebook.ai/developers/tools/elf-opengo
(Emphasis added by me)

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS? :)


This post by Bonobo was liked by: tigerman
Top
 Profile  
 
Offline
 Post subject: Re: Elf OpenGo paper released
Post #19 Posted: Wed Feb 13, 2019 12:34 pm 
Judan

Posts: 5685
Location: Cambridge, UK
Liked others: 321
Was liked: 3007
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
There is indeed a lot more than just that first paper!

- Home page: https://facebook.ai/developers/tools/elf-opengo
- Blog post: https://ai.facebook.com/blog/open-sourc ... -research/
- Online explorer of Elf's view on mistakes, winrate drops etc on large corpus of pro games from history from GoGoD, pick a player, metric, see pretty graphs (note it uses GoGoD's names so Lee Changho is Yi Ch'ang-ho): https://dl.fbaipublicfiles.com/elfopeng ... index.html
- Readme about their pro game analysis: https://dl.fbaipublicfiles.com/elfopeng ... sis/README
- Containing this link to a 3 GB gzip of Elf's analysis of 100k pro games from GoGoD! (Hope JF is ok with this) Big caveat that it assumes 7.5 komi so win% will be off for older games, how valid move suggestions will be is debatable: https://dl.fbaipublicfiles.com/elfopeng ... _sgfs.gzip
- Windows binary (1.3GB!) to play against the real Elf OpenGo rather than LZ converted weights, we can finally see if real Elf suffers from same lack of exploration: https://dl.fbaipublicfiles.com/elfopeng ... ngo_v2.zip


This post by Uberdude was liked by 2 people: Bonobo, Waylon
Top
 Profile  
 
 Post subject: Re: Elf OpenGo paper released
Post #20 Posted: Wed Feb 13, 2019 1:52 pm 
mhlepore wrote:
I hope I don't get in trouble for answering in the wrong place:
not with me :) we need a little light relief with all this heavy stuff about weights and whatnot that makes my brain hurt

mhlepore wrote:
I believe the keima increases black's ability to take sente.
ah, yes, tx, that's what St Mike was saying.

mhlepore wrote:
Which seems to be a thing now.
it's always been a thing - sente is like your car keys, lose it and you have to walk back with a heavy load on your back

mhlepore wrote:
I also don't think it is a given that white is going to butt heads. I've seen O-02 played as well
yes, yes, it's all coming back to me now, except for one thing: it's O2 (or, even better, o2) - you and Billy boy keep your unnecessary Courier Old monotype dashes and leading zeros to yourselves! Before we know it, you will be insisting that all English words are 8 bits long!! Is this what they mean by New World Order?! Big Brother = Bill?!!! omg....

[edit] just had a look at the moves past 43 - it's an ugly ugly game, black seems to be playing suboptimal moves to start kos wherever he can, and some of white's moves look downright simple-minded - so how come they can beat me with one eye closed standing on one leg balancing a bucket of water on their heads?


Last edited by jaca on Wed Feb 13, 2019 2:56 pm, edited 1 time in total.
Top
  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 47 posts ]  Go to page 1, 2, 3  Next

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: Bing [Bot], Google Feedfetcher and 6 guests


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group