It is currently Thu Mar 28, 2024 5:46 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 28 posts ]  Go to page 1, 2  Next
Author Message
Offline
 Post subject: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #1 Posted: Thu Oct 04, 2018 10:09 am 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
I found several interesting things from reviewing a game today in the Korean league between Kim Jiseok 9p (white) and Lee Younggu 9p with Elf v1.

In the game Kim played at 2 with 1 to reach out to his lone k3 stone and prevent black gobbling up the lower side (and then Lee defended the right side at a). Elf thinks it was better for white to invade the right side first with 1. Its first idea as black is to answer with kick at b which it would then tenuki to play at 2: the point seems to be that's a good exchange to make black answer inside now because if you invade after black has a stone at r14 he will answer differently (attach on top or cap or kosumi). So after a few hundred playouts Elf as black decides it should tenuki to 2 on the lower side; interestingly after a few hundred more decides kosumi approach at e3 is better (it has a more severe local followup if ignored so leads to black gobbling up the lower side in exchange for white getting a big solid lower left corner). If black plays 2 what is white's followup on the right side? It shocked me, can you find it?

Click Here To Show Diagram Code
[go]$$Wc Elf suggestion
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X . . . , . . . . . X . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 1 . . |
$$ | . . . . . . . . . . . . . . . . b . . |
$$ | . . . , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . . . . . X . X X X O . . O X . . . |
$$ | . . . . . . . X . . O . O . O X X X . |
$$ | . . . . . . . . . X . X . . . O X O X |
$$ | . . . . . . . . . . . X O O O O X O . |
$$ | . . . O . . . . . , . X O X O X O O . |
$$ | . . . . . 2 . . . O . X O X X X O . . |
$$ | . . . . . . . . . . . . X X X O O O . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]


Hint:
A proverb says my opponent's key point is mine too

Answer:
Click Here To Show Diagram Code
[go]$$Wc Not a typo
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X . . . , . . . . . X . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 1 . . |
$$ | . . . . . . . . . . . . . . . . 3 . . |
$$ | . . . , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . . . . . X . X X X O . . O X . . . |
$$ | . . . . . . . X . . O . O . O X X X . |
$$ | . . . . . . . . . X . X . . . O X O X |
$$ | . . . . . . . . . . . X O O O O X O . |
$$ | . . . O . . . . . , . X O X O X O O . |
$$ | . . . . . 2 . . . O . X O X X X O . . |
$$ | . . . . . . . . . . . . X X X O O O . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]


Here's the game. I didn't record all the way to the end, Kim won by 2.5.


and winrate graph
Attachment:
Kim Lee oct 4 Elfv1 winrate.PNG
Kim Lee oct 4 Elfv1 winrate.PNG [ 324.45 KiB | Viewed 10750 times ]


More to come...


This post by Uberdude was liked by 4 people: Bill Spight, Elom, Gomoto, hyperpape
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #2 Posted: Thu Oct 04, 2018 12:56 pm 
Honinbo

Posts: 9545
Liked others: 1600
Was liked: 1711
KGS: Kirby
Tygem: 커비라고해
Cool. After the not a typo move, S9 seems like a good follow up sometime later. Black isn't totally alive, and white can have a good shape.


By the way, how long does it take to do a "few hundred playouts" on your machine? What kind of hardware do you have? I'm again trying to get an excuse to buy a new computer.

_________________
be immersed

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #3 Posted: Thu Oct 04, 2018 1:39 pm 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Kirby:
Yes, s9 is a good followup if white wants more eyespace and reduce black to even less, but it's not so urgent as even playing first black can't make 2 eyes locally, and Elf sees forcing black to run as valuable. For example if black plays at r10 next (not the best move) then Elf as white would respond with s11 empty triangle to keep black unsettled. My feeling would be to see the 2 white stones as weak and want to jump to safety with p12 and not mind if black lives in gote, but Elf says black is happy to s11 s12 t11 and live (I suppose white is not using m10 ponnuki to full efficiency). With the r11 extend white makes black's q12 attach no longer an effective move to get out whilst leaning on white, so black needs to run with p12 knight move (which white can cut across later) or p11 kosumi.


I did the initial analysis on my work laptop which is not so powerful: about 100 playouts a minute or so. My home PC with a GTX 1060 does 200 in 1 second for Elf. One good thing about being slow is you can see Elf's first feeling (which is more likely to be some fairly sensible shape move I can think of) and how it evolves into more skillful moves with more analysis. This weird :w3: though it wanted to play right from the beginning.

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #4 Posted: Thu Oct 04, 2018 2:05 pm 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
There was an interesting pushing battle in the centre which ended up with a trade for mutual ponnukis, game sequence below.
Click Here To Show Diagram Code
[go]$$Bc Real game
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X . . . , . . . . . X . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . X . . . |
$$ | . . . . . . . 7 6 4 2 9 . . . X . . . |
$$ | . . . . . . . 8 5 3 1 O . . O X . . . |
$$ | . . . . . . . . . . 0 . O . O X X X . |
$$ | . . . . . . . . . . . X . . . O X O X |
$$ | . . . . . . . . . . . X O O O O X O . |
$$ | . . . O . . . . . , . X O X O X O O . |
$$ | . . . . . . . . . O . X O X X X O . . |
$$ | . . . . . . . . . . . . X X X O O W . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]

Click Here To Show Diagram Code
[go]$$Bcm11 Real game cont.
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X . . . , . . . . . X . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . 4 . . . X . . . |
$$ | . . . . . . . X O O O X 2 . . X . . . |
$$ | . . . . . . 5 O X X X O . . O X . . . |
$$ | . . . . . . . 3 . . O . O . O X X X . |
$$ | . . . . . . . . . 1 . X . . . O X O X |
$$ | . . . . . . . . . . . X O O O O X O . |
$$ | . . . O . . . . . , . X O X O X O O . |
$$ | . . . . . . . . . O . X O X X X O . . |
$$ | . . . . . . . . . . . . X X X O O O . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]


What does Elf think? White had just lived in the corner with the marked stone and black moved out to the centre with 1. I thought before living white might force with peep at l2 but is it needed now or can you get it later if black there as reverse sente would be too slow? Elf thinks peep is ok but directly living slightly better. But it does strongly want to play L2 for black 1 and thinks attach lost 12%! An example continuation is hidden below
The direction of play here is quite different which I interpret as follows: black doesn't like white getting l2 peep in sente (but why wasn't it better for white to peep before living in the corner then?) and doesn't want to attach to the white centre group because it sees it as weak and attaching will make it stronger. :w2: stops black making a base easily on lower side. Before 3 Elf actually wanted t2 t2 interlude which I don't understand. 3 then attacks white from the centre whilst developing the right side (which with r12 before reveals as not yet settled so some prophylaxis). This induces running out with the lower side group without strengthening white's centre. 8 being the time to tenuki instead of keep up attack is interesting, if white does attack the nozomi of f5, a shape I'd previously noticed AlphaGo was fond of, crops up as a resource.

Click Here To Show Diagram Code
[go]$$Bc Don't attach to weak group
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X . . . , . . . . . X 8 . . |
$$ | . . X . . . . . . . . . . . . . 9 0 . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . 7 . . . . . . . . |
$$ | . . . , . . . . . , . . 3 . . X . . . |
$$ | . . . . . . . . . . 4 . . . . X . . . |
$$ | . . . . . . . . 6 . . O . . O X . . . |
$$ | . . . . . . . . . . . . O . O X X X . |
$$ | . . . . . . . . 5 . . X . . . O X O X |
$$ | . . . . . . . . . . . X O O O O X O . |
$$ | . . . O . . . . . , . X O X O X O O . |
$$ | . . . . . . 2 . . O . X O X X X O . . |
$$ | . . . . . . . . . . 1 . X X X O O O . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]


White's hane of 2 was only move, but k8 extend -3%, prefers l7 pull back and then expects white f/g3 and then black cuts (if white defended cut instead black is good with f3/g4 area. Elf even thinks g4 directly is better than k8, and white's best response is h4 attach not l7 atari. White's push of 4 should be (-19%) g3 defend. Watching the game live I worried about black's m9 cut in response and Elf says that's best so extend of 5 was slack (-10%). I wonder what was Kim Jiseok's plan if black did cut. Elf doesn't show anything that makes me scared to cut as black. About the best white can do is push out along the N line and then attack the right with q12 but that group can live. 6 push was another overplay, this time Elf wants to defend at g4 slightly over g3. So instead of 7 hane Elf wants to start fighting on the lower side with g4 (white's hane at h8 is no worry, white's best response is shape attack at k6 and then complex fighting develops), but only -3%. Being scared of the m9 cut I thought about n10 for 8 and Elf says it's better than Kim's cut (-7%) but o10 is the way to do that. 9 cut is only move and black is now at 80% (but bear in mind Elf has very strong opinions on slight differences).

How to deal with the cut? It'd be nice if white could capture it. Elf looks at n9 atari initially, then tries Kim's l7 hane (obviously if black m7 cuts it gives white a stones at n8 in sente so then m10 captures the cutting stone), then m7 and after 100k playouts is back to n9 as best. The continuations are similar to before with pushing along n line, invading at r12 and giving up on the 3 stones in the centre for now. Kim's hane is no big deal, -2%, black's answer correct, and then atari ok but slightly (<1%) prefers to exchange m7 for j5 first (damages lower side a bit but I presume Elf is concerned for the safety of the white group and wants n8 to be an eye.

But now comes the big question: why on earth did black not extend from atari with 13? Elf says it's -40%. Is Elf right it's a massive mistake, or is there perhaps an Elf delusion, a ladder it didn't see? Isn't the white group in trouble so why let it off the hook so easily? Maybe Lee thought if he extended and white keeps pushing his right side gets in trouble? Elf suggests a way to avoid that:
Attachment:
kim lee 1.GIF
kim lee 1.GIF [ 169.77 KiB | Viewed 10655 times ]

Note white exchanged m7 for j5 again, if white just n10 directly then after m7 atari black m11 extend is relatively better (but h10 still good) because the white group has less eyespace and black is at 97% instead of 82%. Or maybe Lee thought his atari was more sente (and didn't like the prospect of white getting h7 in sente, is h7 for g8 a good exchange vs white getting h7 if white does end up sacrificing the 3 stones at k9?) and was happy with the resulting ponnuki trade (I can see a possible tewari argument for that). Or maybe they were just playing badly because of short time limits. Maybe Lee thought Kim could have been defensive earlier after the marked black stone and just defended his cut with a handing connection (Elf said o10 best). Then black likewise defends his cut. In this position white should be able to get the peep at a, black's peep at b an be resisted with c so needs preparation at o10 (Elf gives h7 h8 g3 o10 o9 as continuation, note o9 eyespace defence against the peep). But instead of peeping white makes a retarded throw in at h8 and black captures, and black does likewise at m9. Which was worse? My feeling was white's was worse as black's ponnuki faces a more open area so more valuable and white could have got an easy peep there, whilst white's ponnuki has less potential and just makes that group move alive (ignore the lack of l7 k6 exchange for now). However, Elf says before the retarded throw ins white is 40%, but after is at 70%, ie black's was lots worse than white's! I can now see this as white's centre group getting 2 eyes being very valuable to Elf, moreso because it weakenes black's group on the right so white r12 becomes a bigger threat, and black o10 jump no longer threatens an eye destroying peep at n9.
Click Here To Show Diagram Code
[go]$$Wcm8 Tewari of sorts
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X . . . , . . . . . X . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . 1 . c . X . . . |
$$ | . . . . . . . B O O O . b . . X . . . |
$$ | . . . . . . 2 . X X X O . . O X . . . |
$$ | . . . . . . . a . . . . O . O X X X . |
$$ | . . . . . . . . . . . X . . . O X O X |
$$ | . . . . . . . . . . . X O O O O X O . |
$$ | . . . O . . . . . , . X O X O X O O . |
$$ | . . . . . . . . . O . X O X X X O . . |
$$ | . . . . . . . . . . . . X X X O O O . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #5 Posted: Thu Oct 04, 2018 2:20 pm 
Gosei

Posts: 1733
Location: Earth
Liked others: 621
Was liked: 310
R11 is a prototypical example of a practical move :-). It just works.



2 times edited ;-) I need more sleep :-)


Last edited by Gomoto on Sat Oct 06, 2018 8:07 am, edited 2 times in total.

This post by Gomoto was liked by: Elom
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #6 Posted: Thu Oct 04, 2018 4:40 pm 
Gosei

Posts: 1334
Liked others: 0
Was liked: 528
By the way, Lee Younggu just got married with Oh Jeongah.

Image

More pictures:
http://www.cyberoro.com/news/news_view. ... num=524806


This post by trout was liked by 2 people: Uberdude, Waylon
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #7 Posted: Thu Oct 04, 2018 8:25 pm 
Lives in gote

Posts: 388
Liked others: 416
Was liked: 198
Uberdude wrote:
If black plays 2 what is white's followup on the right side? It shocked me, can you find it?


Wow, absolutely incredible... Shocking is an understatement...

_________________
Sorin - 361points.com

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #8 Posted: Fri Oct 05, 2018 12:40 am 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
So far all of Elf's suggestions have seemed plausible (even the bizarre :w3: with some thought) and I can believe they are better moves than what the pros played in the game. Of course it's possible the pros had some good response that Elf didn't see (particularly involving ladders), but when it comes to judgement of positions, particularly quiet ones or even fighting without low liberties I tend to think Elf is better than the pros. However, here's a big counterexample. White has just extended at 1 which completes a big centre territory. It also threatens to kill the black corner with a-e, a not particularly hard tsumego that I expect even ddks can find (the corner eyespace is common and easy, the throw-in to falisfy eye maybe some 10k would overlook in a game but I expect most would get in a problem). But Elf is seemingly oblivious to this and wants to start push and cut on the top side. To start with it thinks :b2: has 63% win though by 30k playouts this has dropped to 55%. It also likes h12. It doesn't even consider making life at e5 as Lee did.

Click Here To Show Diagram Code
[go]$$Wc Elf's plan as black
$$ ---------------------------------------
$$ | . . . . 7 . . . . . . . . . . . . . . |
$$ | . a . 5 4 3 6 8 0 . . . . . . . . . . |
$$ | . X . O O 2 O . 9 . . . . . . . . . . |
$$ | . . O , X X O . . , . . . . . X . . . |
$$ | . X X X . X X O . . . . . . . . . . . |
$$ | . . . O O . X X . . . . . . . . X . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . O . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . 1 . . . X . X . . O . . O X . . . |
$$ | . . O X . . . X O O O . O . O X X X . |
$$ | . O X O O . . O O X . X . . . O X O X |
$$ | . O X X . O O O X X X X O O O O X O . |
$$ | . X . e X O X X O X . X O X O X O O . |
$$ | . X X X O O X O O O X X O X X X O . . |
$$ | . . d c . a X X O b O X X X X O O O . |
$$ | . . . . . . . X X O . . . O . X . . . |
$$ ---------------------------------------[/go]


Maybe it's looking at the big picture and saying the corner is small: if white kills it and black gets 2 moves on top black is good? If I play white 1-3 and kill the corner Elf would play 4 at 10k playouts with 56% win. But that's based on it thinking white will play 5 as another move in the corner, but that's unnecessary as black is already dead!

Click Here To Show Diagram Code
[go]$$Wc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . 4 . . . . . . . . . . . . . |
$$ | . X . O O B O . . . . . . . . . . . . |
$$ | . . O , X X O . . , . . . . . X . . . |
$$ | . X X X . X X O . . . . . . . . . . . |
$$ | . . . O O . X X . . . . . . . . X . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . O . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . O . . . X . X . . O . . O X . . . |
$$ | . . O X . . . X O O O . O . O X X X . |
$$ | . O X O O . . O O X . X . . . O X O X |
$$ | . O X X . O O O X X X X O O O O X O . |
$$ | . X . , X O X X . X . X O X O X O O . |
$$ | . X X X O O X . . . X X O X X X O . . |
$$ | . . . 3 . 1 X X . 2 O X X X X O O O . |
$$ | . 5 . . . . . X X O . . . O . X . . . |
$$ ---------------------------------------[/go]


So if I play 4 in the corner to show Elf black is dead and play out the death of the corner (losing ko threats for black but otherwise should cause no change in judgement of board) Elf realises black is terrible and only 1% win. Its preferred move is the b7 cut, the "maybe they tenuki my sente move and let me live" silly sente desperate bots are known for when losing. Playing the g18 hane at the top is simiarly bad 1.5%. So Elf just totally misunderstood this simple life and death problem. So how much can we trust its previous judgements? IT does at least say the d7 hane was a big mistake (should b7) and that would avoid white being able to close the centre with the sente threat for life on the corner due to black's self-inflicted shortage of liberties. Also instead of h14 it sees a sharp tactical attachment at c11 so if Elf was playing black it wouldn't have got itself into this position where it could fail a kyu tsumego, but that it does fail so spectacularly when presented with it is surprising and disappointing.
Attachment:
kim lee 2.GIF
kim lee 2.GIF [ 229.27 KiB | Viewed 10593 times ]


This post by Uberdude was liked by 2 people: Bill Spight, hyperpape
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #9 Posted: Fri Oct 05, 2018 1:40 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Uberdude wrote:
So Elf just totally misunderstood this simple life and death problem. So how much can we trust its previous judgements? IT does at least say the d7 hane was a big mistake (should b7) and that would avoid white being able to close the centre with the sente threat for life on the corner due to black's self-inflicted shortage of liberties. Also instead of h14 it sees a sharp tactical attachment at c11 so if Elf was playing black it wouldn't have got itself into this position where it could fail a kyu tsumego, but that it does fail so spectacularly when presented with it is surprising and disappointing.


The strength of current top bots lies in global evaluation and search. The strength of humans lies in local evaluation and search, and in global depth first search of one lane roads, such as ladders. There are things, such as the simple L&D position in the bottom left corner, that the bots have not learned by self play because they seldom get into those positions or both sides play incorrectly when they do. Given infinite time they will catch up to humans in such situations and surpass them. But I actually think that will happen fairly soon, because human programmers will either figure out different ways to program the bots or different ways to train them. :)

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

Visualize whirled peas.

Everything with love. Stay safe.


Last edited by Bill Spight on Fri Oct 05, 2018 1:52 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #10 Posted: Fri Oct 05, 2018 1:49 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
BTW, what if you given Elf this position with Black to play?

Click Here To Show Diagram Code
[go]$$Bc Easier corner
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . X . O O . O . . . . . . . . . . . . |
$$ | . . O , X X O . . , . . . . . X . . . |
$$ | . X X X . X X O . . . . . . . . . . . |
$$ | . . . O O . X X . . . . . . . . X . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . O . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . W . . . X . X . . O . . O X . . . |
$$ | . . O X . . . X O O O . O . O X X X . |
$$ | . O X O O . . O O X . X . . . O X O X |
$$ | . O X X . O O O X X X X O O O O X O . |
$$ | . X B . X O X X O X . X O X O X O O . |
$$ | . X X X O O X O O O X X O X X X O . . |
$$ | . . . . . . X X O . O X X X X O O O . |
$$ | . . . . . . . X X O . . . O . X . . . |
$$ ---------------------------------------[/go]

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

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #11 Posted: Fri Oct 05, 2018 2:16 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
Quote:
The strength of current top bots lies in global evaluation and search. The strength of humans lies in local evaluation and search, and in global depth first search of one lane roads, such as ladders.


I'm unclear as to whether global and local here apply just to 'evaluation' or to both 'evaluation and search'.

The reason I raise it is that I'm also unclear how Elf manages to see the 'shocking' R11 narabi earlier on. It doesn't seem like the sort of move it would see very often even with millions of self-training games, but even if it did have this as a globally good candidate move on the basis of self training, it would still need to be good at local search to justify this move (and maybe good enough to find a long ladder?). Or have (a) I completely misunderstood how Elf finds it moves or (b) wrongly assumed it has chosen the narabi on the basis of a search whereas it was chosen solely on the grounds of global evaluation?


This post by John Fairbairn was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #12 Posted: Fri Oct 05, 2018 2:31 am 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
FWIW, LeelaZero #157 also finds r11 quickly, also finds e3 over f3 for black after r12 (but needed more playouts than Elf), also thinks black should have extended from the atari, and also fails the bottom left tsumego.


This post by Uberdude was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #13 Posted: Fri Oct 05, 2018 2:57 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
John Fairbairn wrote:
I'm unclear as to whether global and local here apply just to 'evaluation' or to both 'evaluation and search'.

The reason I raise it is that I'm also unclear how Elf manages to see the 'shocking' R11 narabi earlier on. It doesn't seem like the sort of move it would see very often even with millions of self-training games, but even if it did have this as a globally good candidate move on the basis of self training, it would still need to be good at local search to justify this move (and maybe good enough to find a long ladder?). Or have (a) I completely misunderstood how Elf finds it moves or (b) wrongly assumed it has chosen the narabi on the basis of a search whereas it was chosen solely on the grounds of global evaluation?


My guess is that it has to do with history and path dependency. That is, in the early stages of self play the games are almost random. In that environment the narabi is fairly powerful. (So is the one point eye, but that usually takes four stones instead of two.) So I think that the narabi is a shape that the zero bots would discover early in their development. I also think that the raw attachment and diagonal approach would be learned early on. So these local plays will often be suggested as possible plays when building the search tree, but the search is still global. Anyway, that's my guess.

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

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #14 Posted: Fri Oct 05, 2018 3:48 am 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Seeing as the f17 push and double cut is a locally hot shape, moreso given black's nor white's group there is settled, I tried adding some vaguely sensible (but better for white: j15 atari, f17 connect, c18 slack kosumi, k17 good extend, h16 capture slack, h17 simple block, o17 slack enclosure (could be m17), m17 defend extend, b16 super slack because I need it to be white sente) equal number of stones to the top side to calm it down and see if without that distraction Elf would find the kill. It did after a few hundred playouts as shown below (and the principal variation is playing out the kill including b7 cut silly sente), but started off wanting to play q18. Plus as white is 96% maybe it's not such a good test.
Attachment:
kim lee 3.PNG
kim lee 3.PNG [ 1018.78 KiB | Viewed 10547 times ]


Here's a closer position. To start with white wanted to 3-3 at r17 with ~62%, but around 100 playouts it discovered the lower left kill and sees that's much better at 87%.
Attachment:
kim lee 4.PNG
kim lee 4.PNG [ 1012.72 KiB | Viewed 10547 times ]


So Elf can find the kill as white fairly quickly if there aren't other urgent things too on the board to distract it. But it doesn't seem so good as black at seeing the impending death: if in the last diagram white plays a rubbish gote like d19 black doesn't want to live at e5 but play q13 or o12 or f13 (at 500 playouts, I'll do more).

Update: Did the: "can black see it needs to live after white d19" test to 50k playouts. For the first 10k or so it was looking at centre moves like h12, o12, f13 etc in the 70s. It did give 26 of 50000 to e5 but with only 71%. At 50k it seems to have settled on a rather curios idea at 75.5%: play t3 which is just a sente move for life of that corner, it expects white to tenuki to kill the lower left, then black can kill the (roughly equally sized) lower right and then it expects white to waste another move killing the already dead lower left and black continues with reducing in the centre. But if I take the part of white and just answer t3 at t2 to decline the trade (because I think Elf is stupid and I can kill lower left later) then Elf goes back to wanting to play in the centre rather than living at e5. This seems to be an instance of the bot habit of playing pointless sentes which just pushes potential bad things in the future off the radar, I believe called the "horizon effect".


This post by Uberdude was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #15 Posted: Fri Oct 05, 2018 6:27 am 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
Uberdude wrote:
So Elf can find the kill as white fairly quickly if there aren't other urgent things too on the board to distract it. But it doesn't seem so good as black at seeing the impending death.


Friedrich Nietzsche wrote:
Menschliches, Allzumenschliches

;)

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

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #16 Posted: Fri Oct 05, 2018 7:28 am 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Bill Spight wrote:
BTW, what if you give Elf this position with Black to play?
Click Here To Show Diagram Code
[go]$$Bc Easier corner
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . X . O O . O . . . . . . . . . . . . |
$$ | . . O , X X O . . , . . . . . X . . . |
$$ | . X X X . X X O . . . . . . . . . . . |
$$ | . . . O O . X X . . . . . . . . X . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . O . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . W . . . X . X . . O . . O X . . . |
$$ | . . O X . . . X O O O . O . O X X X . |
$$ | . O X O O . . O O X . X . . . O X O X |
$$ | . O X X . O O O X X X X O O O O X O . |
$$ | . X B . X O X X O X . X O X O X O O . |
$$ | . X X X O O X O O O X X O X X X O . . |
$$ | . . . . . . X X O . O X X X X O O O . |
$$ | . . . . . . . X X O . . . O . X . . . |
$$ ---------------------------------------[/go]


It still wants to f17 push and still doesn't consider e5 life, but things have changed a bit: it only gives f17 45% win (and falling at 1k playouts) because now it does expect in the principal variation white to kill the corner instead of answering the push, but like before expects white to spend 2 gotes to kill it when only 1 is needed so black gets 3 moves (instead of the 2 he deserves) on the top side. However, if you then play it out then after f17, kill corner, g18 h18 f18 (actually f18 can cut) then it realises white doesn't need another move to kill the corner and can play m17 for 83%.

P.S. black h18 as a followup to f17 instead of normal f18 or g18 is interesting.


This post by Uberdude was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #17 Posted: Fri Oct 05, 2018 9:09 am 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Black to play, I didn't mark white's last move to avoid that influencing you:
Click Here To Show Diagram Code
[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O O . O . . . . . . . . . . . . |
$$ | . . . , X X O . . , . . . . . X . . . |
$$ | . . X . . . X O . . . . . . . . . . . |
$$ | . . . . . . X . . . . . . . . . X . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O , . . . . . , . O . . . X . . . |
$$ | . . . . . . . X O O O . O . . X . . . |
$$ | . . . . . . X O . . . O . . O X . . . |
$$ | . . . . . . . X O O O . O . O X X X . |
$$ | . . X . . . . O . X . X . . . O X O X |
$$ | . . . . . . . . X O . X O O O O X O . |
$$ | . X O O . . . . . X . X O X O X O O . |
$$ | . . X X . O . . X O . X O X X X O . . |
$$ | . . . . . . . . . . O X X X X O O O . |
$$ | . . . . . . . . . . . . . O . X . . . |
$$ ---------------------------------------[/go]

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #18 Posted: Sat Oct 06, 2018 4:28 am 
Lives in sente

Posts: 827
Location: UK
Liked others: 568
Was liked: 84
Rank: OGS 9kyu
Universal go server handle: WindnWater, Elom
Quote:
If black plays 2 what is white's followup on the right side? It shocked me, can you find it?


Since it is surprising, P11 as the white mid-lower group helps :w1: despite not looking like it does, and the ladder works

Quote:
Hint: A proverb says my opponent's key point is mine too

...Q12...?.

Quote:
Answer: Not a typo


Okay. Wow.

_________________
On Go proverbs:
"A fine Gotation is a diamond in the hand of a dan of wit and a pebble in the hand of a kyu" —Joseph Raux misquoted.


Last edited by Elom on Sat Oct 06, 2018 6:08 am, edited 1 time in total.
Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #19 Posted: Sat Oct 06, 2018 5:57 am 
Lives in sente

Posts: 827
Location: UK
Liked others: 568
Was liked: 84
Rank: OGS 9kyu
Universal go server handle: WindnWater, Elom
sorin wrote:
Uberdude wrote:
If black plays 2 what is white's followup on the right side? It shocked me, can you find it?


Wow, absolutely incredible... Shocking is an understatement...


Saying you only see moves like it in problem books is not enough. It's a move for movies.

_________________
On Go proverbs:
"A fine Gotation is a diamond in the hand of a dan of wit and a pebble in the hand of a kyu" —Joseph Raux misquoted.

Top
 Profile  
 
Offline
 Post subject: Re: Elfv1 highlights of Kim Jiseok vs Lee Younggu
Post #20 Posted: Sat Oct 06, 2018 6:03 am 
Lives in gote

Posts: 311
Liked others: 0
Was liked: 45
Rank: 2d
John Fairbairn wrote:
Quote:
The strength of current top bots lies in global evaluation and search. The strength of humans lies in local evaluation and search, and in global depth first search of one lane roads, such as ladders.
I'm unclear as to whether global and local here apply just to 'evaluation' or to both 'evaluation and search'.
I'm not sure how it was meant, but fairly sure it applies to both. In fact I would stress global search more than global evaluation (bot evaluation itself is not necessarily better than pros, even Alphago was top amateur only without search). But global evaluation AND (more or less / reasonably) correct global search is something humans don't have an answer for.

Quote:
The reason I raise it is that I'm also unclear how Elf manages to see the 'shocking' R11 narabi earlier on. It doesn't seem like the sort of move it would see very often even with millions of self-training games, but even if it did have this as a globally good candidate move on the basis of self training, it would still need to be good at local search to justify this move (and maybe good enough to find a long ladder?). Or have (a) I completely misunderstood how Elf finds it moves or (b) wrongly assumed it has chosen the narabi on the basis of a search whereas it was chosen solely on the grounds of global evaluation?
Usually it only searches in moves chosen by global evaluation / policy (which is why it still has blind spots). But IMO it doesn't justify candidates - it refutes them. Moves that drop winrate are phased out, what's left is played. It's much harder to phase in new candidates as that takes much more search effort.

This particular case may also have to do with generalizability: some things generalize better, so easier to learn even from different positions, others may need more specialized (so more rare) occurrences.

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 28 posts ]  Go to page 1, 2  Next

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