It is currently Mon Apr 12, 2021 5:27 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 16 posts ] 
Author Message
Offline
 Post subject: Mysterious new bot 'define' beats Zen to top CGOS.
Post #1 Posted: Sat Dec 02, 2017 1:14 am 
Judan

Posts: 6676
Location: Cambridge, UK
Liked others: 432
Was liked: 3662
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
A new bot called define is doing very well on the computer go server. It's top of the regular elo rating list (not enough games quite yet for the daily Bayes Elo), beat the top Zen version 15.5 2c1g 2-0, Zen 15.6 2c1g 7-1 and overall 89-5. Its play is reminiscent of AlphaGo with early 3-3s, peeping the resulting wall, liking attach not slide after 4-4 approach. Here's a few sample games.

Vs top rated 4122 Zen 15.5. The 8-10-12 combo was AlphaGo-esque, and the armpit hit invasion of 38 particularly so.


Another vs Zen 15.5. Early 3-3 invasion and then wedge same distance Redmond noticed AlphaGo likes and then peep wall and sabaki. I thought it was in trouble in bottom left fighting but got out with a ko.


Vs rather weaker 2890-rated Leela with nice kill top left after which seems over. Maybe Leela should take sente from 3-3 invasion to pincer left side? Reponse to top side shoulder hit also clumsy.


Here's the one loss against Zen 15.6. Why didn't it push one more to cleanly live top right? Seems like it didn't do so great in left fighting and then missed cut for semeai it lost and then allowed 2 eyes and went bonkers (1-1 point!).


This post by Uberdude was liked by 4 people: Bonobo, dhu163, ez4u, Joaz Banbeck
Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #2 Posted: Sat Dec 02, 2017 1:21 am 
Lives in sente

Posts: 727
Liked others: 44
Was liked: 218
GD Posts: 10
I somehow link 'define' with 'just as expected' which is another name of FineArt.
But then again, without knowing hardware there's nothing to be excited about.
And Tianrang also play there as well as training for the soon AI ryusei cup next week. I might create a thread on this event which will Livestream by Nico video

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #3 Posted: Sat Dec 02, 2017 6:51 am 
Lives in sente

Posts: 975
Liked others: 0
Was liked: 167
pookpooi wrote:
........
But then again, without knowing hardware there's nothing to be excited about.
.........


VERY important. Without knowing what the program is, for all you know could be the SAME program running with more resources << I am not saying that this is the situation, just putting it in here to point out the importance of hardware. Until a program is at the point of diminishing returns >>

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #4 Posted: Sat Dec 02, 2017 9:10 am 
Lives in gote

Posts: 646
Location: Austin, Texas, USA
Liked others: 52
Was liked: 216
Jeez you guys are downers. Why not get excited? This is the first time any bot has even come close to Zen on CGOS, let alone topping it in the ratings. On BayesElo the top 12 bots are all Zen. So I think it is huge news that something else is top of the live Elo list and will probably be top of the BayesElo list tomorrow.

Is it really necessary to point out in every thread on computer go that hardware matters? Isn't that obvious?


This post by yoyoma was liked by 3 people: Bonobo, Marcel Grünauer, Uberdude
Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #5 Posted: Sat Dec 02, 2017 10:10 am 
Dies in gote

Posts: 36
Liked others: 0
Was liked: 6
Rank: Europe 5 dan
KGS: Flashgoe
It wouldn't hurt those guys behind define go program to mention hardware as many programs on CGOS did. To not confuse spectators.

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #6 Posted: Sat Dec 02, 2017 10:26 am 
Judan

Posts: 6676
Location: Cambridge, UK
Liked others: 432
Was liked: 3662
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Yoyoma, thankfully I have not installed pookpooi's excitement regulator, and I've learnt to treat Mike's broken record as amusing rather than annoying.

It's not just the amount of wins that interest me, but their manner. Did define independently find these Ag-style moves, was it trained on the AG games (small data set but could you give it a big weight?), or some hand coding?

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #7 Posted: Sat Dec 02, 2017 12:55 pm 
Lives in sente

Posts: 727
Liked others: 44
Was liked: 218
GD Posts: 10
yoyoma wrote:
will probably be top of the BayesElo list tomorrow.

Look like it stop at 94 games.
It's driving me crazy, as someone who see CGOS everyday, I saw so many programs that stop playing just a few game before it's 100th game, thus preventing it from appear in Bayes elo page. Yes, I'm pointing at you, Leela! (that ran by many users)

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #8 Posted: Sat Dec 02, 2017 4:34 pm 
Judan

Posts: 6676
Location: Cambridge, UK
Liked others: 432
Was liked: 3662
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Something I wondered replaying the first game I posted is how would define handle the knight move instead of kick response to the early low approach inside mini Chinese (Chen Yaoye 9p suggested this when the humans kept losing with the kick against Master coming low into their Chinese formation). This game against PARANOiA (a strong bot just behind Zen I've not heard of before) answers that (immediately live in corner), plus interesting fighting later. Note also Black played the modern human joseki top left.

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #9 Posted: Sun Dec 03, 2017 12:32 am 
Lives with ko

Posts: 135
Liked others: 40
Was liked: 21
Rank: KGS 6 dan
New opening theories have just been made here on L19. I totaly love this keima response in mini chinese. Unlinke kosumi-tsuke which seems to be what white wants.

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #10 Posted: Sun Dec 03, 2017 7:06 pm 
Tengen

Posts: 4347
Location: North Carolina
Liked others: 489
Was liked: 721
Rank: AGA 3k
GD Posts: 65
OGS: Hyperpape 4k
Mike Novack wrote:
Without knowing what the program is, for all you know could be the SAME program running with more resources
Seems less likely if the style changes dramatically. The changes in Alpha Go's style corresponded with different techniques for building the underlying network, not adding more compute.

_________________
Occupy Babel!

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #11 Posted: Sun Dec 03, 2017 10:26 pm 
Lives in sente

Posts: 727
Liked others: 44
Was liked: 218
GD Posts: 10
hyperpape wrote:
Seems less likely if the style changes dramatically. The changes in Alpha Go's style corresponded with different techniques for building the underlying network, not adding more compute.

You might be correct but we actually lack evidence, the game records of the same version of AlphaGo running on different hardware.

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #12 Posted: Mon Dec 04, 2017 4:40 pm 
Lives in gote

Posts: 430
Location: Adelaide, South Australia
Liked others: 173
Was liked: 224
Rank: Australian 2 dan
GD Posts: 200
Bohdan wrote:
It wouldn't hurt those guys behind define go program to mention hardware as many programs on CGOS did.


Where do you get info about hardware for the programs running on CGOS? I can't see anything from browsing the CGOS web site.

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #13 Posted: Tue Dec 05, 2017 5:13 am 
Lives in gote

Posts: 665
Liked others: 292
Was liked: 329
Uberdude wrote:
Note also Black played the modern human joseki top left.


About move 75 (the atari at e15): Wouldn't it have been better to first play the hane at b17? Then White couldn't have blocked with a counter-hane. After the atari and connect, White can block, as in the game at moves 81 and 84.

Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #14 Posted: Tue Dec 05, 2017 6:56 am 
Judan

Posts: 6676
Location: Cambridge, UK
Liked others: 432
Was liked: 3662
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Marcel Grünauer wrote:
Uberdude wrote:
Note also Black played the modern human joseki top left.


About move 75 (the atari at e15): Wouldn't it have been better to first play the hane at b17? Then White couldn't have blocked with a counter-hane. After the atari and connect, White can block, as in the game at moves 81 and 84.


I'm not strong / studied the game enough to say something is a mistake with confidence, but this is a standard conundrum with this shape: both players are somewhat reluctant to choose a local move because it commits them so makes it more obvious for the opponent what forcing move to play or not. Often in human games white will play 32 at e15, and then black will b17 hane and white will extend at c18 so white gets a strong shape with e15 poking to centre but black b17 in sente makes his group almost alive locally which is important for later fighting. On the other hand if white descends at b17 then black's atari at e15 is a joy to play and although it is possible to tenuki white would probably connect to stop black sente point and eyeshape gain and preserve f14 peep, but then b17 does become inefficient (you wouldn't play it if black atarid first). Sometimes white will just connect at d17 (so can block b17 or jump to f14 later) but it's a little slow, though I have some fuzzy memory AlphaGo played this in a self-play game.

So with that intro, I think that explains why white tenukid and cut from lower left to start a big fight. Now to Marcel's question, wouldn't 75 at b17 be better? The key issue at present is black is attacking white's loose group of e11 h12 etc stones. The black group below is not alive either, nor the one above, but as black could cut at d11/12 to kill that one side stone and connect the two groups neither is in big trouble. In fact that being unplayed (either black doing it or white defending with e.g. e12) for so long puzzles me a bit, but I think the explanation is black judges it as too slow and wants to attack the whole group, because if black attacks from outside white can't go back to defend because it's heavy and black would surround more and may kill. However, the threat of e12 does carry some power for white as it could then set up attacks on the black groups both above and below, so there's a bit of a game of chicken going on. In fact later in the game white sacrificed the whole tail so a move like e12 would have just become wasted, as would a black cut at d11.

So if black did play b17 for move 75 what would white do? One idea I have is white might actually hane at b18 as a sacrifice as if black ataris to capture the corner white can then jump to f14 which somewhat connects up to the h12 stone (which is what black prevented with the game line). Certainly white wouldn't just pull back to c18 as then the e15 atari is great and white can't connect without looking stupid. Maybe e15 for white? but then c18 atari and if white doesn't connect then e16 aji is bad, and if you do connect e15 was better at f14 so I prefer the b18 sacrifice. White could also d17 connect which makes miai of b18 and f14, but the follow up for white if black tenukis is less point-valuable (c.f. if b18 and tenuki then you can b16 capture) but on the plus side black lost the e15 atari. Or just tenuki to the centre fight maybe?


This post by Uberdude was liked by: Marcel Grünauer
Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #15 Posted: Tue Dec 05, 2017 8:48 am 
Honinbo

Posts: 10673
Liked others: 3593
Was liked: 3339
Marcel Grünauer wrote:
Uberdude wrote:
Note also Black played the modern human joseki top left.


About move 75 (the atari at e15): Wouldn't it have been better to first play the hane at b17? Then White couldn't have blocked with a counter-hane. After the atari and connect, White can block, as in the game at moves 81 and 84.


Nor am I able to say what is best, but here is an observation. :) The hane is stronger in the corner, but Black focuses on the fight in the center, the result of which is a strong wall there and eventually a large territory with the capture of a few White stones. For humans this is a question of judgement.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.


This post by Bill Spight was liked by: Marcel Grünauer
Top
 Profile  
 
Offline
 Post subject: Re: Mysterious new bot 'define' beats Zen to top CGOS.
Post #16 Posted: Tue Dec 05, 2017 10:37 am 
Lives in gote

Posts: 665
Liked others: 292
Was liked: 329
Uberdude wrote:
I'm not strong / studied the game enough to say something is a mistake with confidence, but this is a standard conundrum with this shape


Thank you for this very detailed analysis!

It reinforces my impression that "joseki moves" are no more than locally acceptable sequences but that recognizing and judging the many subtle kinds of aji that are left in a position and that are intertwined with the stability of surrounding groups is what is required to be a strong player.


This post by Marcel Grünauer was liked by: johnsmith
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 16 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


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:  
cron
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group