| Life In 19x19 http://www.lifein19x19.com/ |
|
| Opening problems for AI: Problem 3 http://www.lifein19x19.com/viewtopic.php?f=15&t=17558 |
Page 1 of 1 |
| Author: | Bill Spight [ Sat Jun 13, 2020 10:06 pm ] |
| Post subject: | Opening problems for AI: Problem 3 |
Requiring the human play to be a likely mistake is a bit of a challenge. Even if a pro 9 dan makes the play, how sure can I be that other pros would make the same play? In any event, in this position a pro 5 dan made a play that follows a general principle of the fuseki, and I would have made the same play. 
|
|
| Author: | Bill Spight [ Sat Jun 13, 2020 10:51 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
| Author: | ez4u [ Sat Jun 13, 2020 11:14 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
I have not looked at the hide yet. My initial reaction is this does not look like we are still in the opening. We'll see! My real reason for posting is to ask you not to go too fast in introducing your problems. I find the larger question of coming to grips with what it is possible to learn from the AI's fascinating. However, these days I do not have either the personal time or the hardware to motor through several of these per day! I can't be the only one either. So gently, gently good sir! Please? 
|
|
| Author: | Bill Spight [ Sun Jun 14, 2020 12:33 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
ez4u wrote: I have not looked at the hide yet. My initial reaction is this does not look like we are still in the opening. We'll see! Well, it's borderline, but the human play is in accordance with an opening principle. It was inspired by one of Fujisawa Hideyuki's problems, which is also late in the opening. Quote: My real reason for posting is to ask you not to go too fast in introducing your problems. I find the larger question of coming to grips with what it is possible to learn from the AI's fascinating. However, these days I do not have either the personal time or the hardware to motor through several of these per day! I can't be the only one either. So gently, gently good sir! Please? I must say I have been impressed with your observations and discussion of these matters. It's a real challenge to learn from black boxes.
|
|
| Author: | Kirby [ Sun Jun 14, 2020 4:30 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
| Author: | Harleqin [ Sun Jun 14, 2020 5:49 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
My instinct would be the one-space high shimari in the top left. I would not like to approach the black thickness too early, and the shimari looks bigger to me. Why not low? Because then Black plays on its shoulder, which seems ideal to me. Why not far? Seems too thin in view of the black thickness in the centre. |
|
| Author: | Tryss [ Sun Jun 14, 2020 7:18 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
My Katago (20b of the end of april run) want this : |
|
| Author: | Bill Spight [ Sun Jun 14, 2020 12:14 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
Just to show you that the principle is sound: From GoGoD 1993-04-03d, Takemiya Masaaki, 9 dan (W) vs. Yang Hui, 8 dan. And this: From GoGoD 2001-04-08c, Rin Kaiho, 9 dan (W) vs. O Meien, 9 dan.  is not Elf's top choice, a is. However, the estimated loss of by comparison is only 0.1%. All same same.BTW, not only follows the principle of limiting White's expansion of its sphere of influence, it also has a good follow-up which has a good follow-up, which has a good follow-up.  What's not to love?
|
|
| Author: | Uberdude [ Sun Jun 14, 2020 12:59 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
I suppose in the examples Elf agrees the side extension limits a moyo that needs limiting because it is thick and big, whereas in the first position Black's top side is nothing to be scared of so you shouldn't play locally bad move for non applicable global reasons. |
|
| Author: | Bill Spight [ Sun Jun 14, 2020 1:08 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
Uberdude wrote: I suppose in the examples Elf agrees the side extension limits a moyo that needs limiting because it is thick and big, whereas in the first position Black's top side is nothing to be scared of so you shouldn't play locally bad move for non applicable global reasons. Most players would, I think, have no trouble making such extensions from the 4-4 point instead of the 3-4 point. The question of when extending from the 3-4 point is good or not is, I am afraid, fairly advanced. |
|
| Author: | sorin [ Sun Jun 14, 2020 2:52 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
Bill Spight wrote: Uberdude wrote: I suppose in the examples Elf agrees the side extension limits a moyo that needs limiting because it is thick and big, whereas in the first position Black's top side is nothing to be scared of so you shouldn't play locally bad move for non applicable global reasons. Most players would, I think, have no trouble making such extensions from the 4-4 point instead of the 3-4 point. The question of when extending from the 3-4 point is good or not is, I am afraid, fairly advanced. Indeed, it takes mental effort to discard the normal follow-up from an empty 3-4 point, which is a shimari, and instead extend in "the wrong local direction". For humans I mean. I wonder whether there is any equivalent effort for AI, in terms of policy net / value net tension
|
|
| Author: | lightvector [ Sun Jun 14, 2020 3:11 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
KataGo recent stronger unreleased 40 blocks, lots of playouts, analysisWideRootNoise = 0.1 to force evaluation of lots of options. Winrates from black's perspective. And one of the subsequent examples: |
|
| Author: | ez4u [ Sun Jun 14, 2020 8:08 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
IMHO the point to be taken here is that in the original position White overvalued the top. Personally I would not have chosen the game play because the top is relatively narrow. The left side looks wider. I am too lazy to do much more than apply O Meien's dictum, "Play on the wider side." However, the bots would find that equally facile. I have to get my pretty picture into the mix like others so here is the current 40b after 1 million playouts. Note that I added the forcing play in the lower right corner. In the end blue is the continuation in the lower right. This is the same as lightvector's post with the newer net (note that I used komi = 6.5 so the winrates/scores are different). However, along the way to the final result, the calculations are throwing up basically three alternatives: enclose the top left, enclose the bottom left, seal off the bottom right. These three areas are interrelated. As the program tests different alternatives, we can see the same things are expected to occur. 1. If White encloses the top left, Black will invade the bottom left. The alternatives for enclosing the top left are tested/evaluated based on how White can respond to a Black invasion of the bottom rather than the effect on the top side. 2. If White encloses the bottom left, Black will play in the top left. However, an approach from along the top is not a real choice. We can look at some of the analysis, but basically Black will not get enough if White secures the upper left corner and side in exchange for the top, having already played in the bottom left. So instead, normally Black will attach against the White stone at the top by playing on the star point. The question then becomes who ends in sente in order to continue at the bottom right. 3. If White continues in the bottom right, Black will not answer locally. Below is the result of leaving the box running all night after White plays in the bottom right (2.5 million playouts!  ). Note that Black will no longer invade the bottom left but rather approach along the left side. The choice of continuing on the right for White is based in part on changing the values in the bottom left.These are the interactions that katago is using to evaluate the choices for White. In the initial position, whether to play along the top or enclose the top left corner would have more to do with the value of the bottom right than the value of the top! This breadth of "vision" is one of the main points of bot analysis. Can mere mortals learn to emulate it? "To see or not to see, that is the question!" 
|
|
| Author: | John Fairbairn [ Mon Jun 15, 2020 3:10 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
In threads of this type there is constant mention of the notion that bots prefer corners to the sides, to an extent that perhaps many of us here haven't quite got attuned to. I was therefore brought up sharp by a headline on Shibano Toramaru's latest side trip on his 15-month journey through the "Fuseki Revolution": "Speed over bad aji in the corner". He was showing a preference for a side move over a corner move! Bear in mind that Shibano's way of working is not to sit in front of a computer playing with Leela or whatever. He spends all his time on the internet looking at what other pros are playing as a result of their AI research. He absorbs the new moves and thinks for himself about why they have replaced the old moves. Others do the grunt work. He would make a great entrepreneur! Just in passing, to explain the reason for my special fascination with Shibano, most of you will be familiar with Michael Redmond's superb ability to distil complex flavours into a spoonful of amber nectar that even we amateurs can appreciate with delight. Shibano has exactly the same ability but adds to that the insight of a Meijin. And with his series now into its 69th episode, the quaich is overflowing! An important point is that Shibano is not recommending moves himself. He is showing what a majority of other pros are doing. We are getting a large body of opinion, which makes his insights especially useful. The wider context of what he was saying here is making an approach against a corner and leaving it, as White has done in the lower left. At some point, though, White has to handle his deserted stone, and the triangled move is apparently still a popular way of doing that. If Black makes a kamae (construction) move at A, White will handle himself easily with a sequence starting at B, which not only strengthens himself on the outside but exposes a bit of aji in the corner. If Black plays forcing moves on both sides of the White sgian dubh he will end up just honing the blade and so playing thank you moves for White, the point there being that with the weakness at C Black's stones are not really working in unison and he will end up with bleeding fingers. Hitherto, the normal move in this situation has therefore been Black D - a honte. This eliminates the bad aji in the corner. But, since AI has been on the scene, this way of playing has been disappearing among pros. The reason, says Shibano, is that White will jump to F. Given that extra White strength, a pro as Black would then feel obliged to defend at B, whereupon White would play the tobisagari at G. And even we now have a strong inkling that this is the sort of centre-dominating shape that bots live. So, what has been happening is that Black players are now choosing the faster-paced E to take control of the side. This leaves bad aji in the corner. White can live there. But if White does take that option, Black seals him in, and the Black stone at G turns out to be an absolutely tight keystone, turning the White sgian dubh into a blunt and rusty pen-knife. Sides before corners, eh? The plot thickens! |
|
| Author: | Knotwilg [ Mon Jun 15, 2020 3:47 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
Late to the party: |
|
| Author: | Uberdude [ Fri Jun 19, 2020 12:42 pm ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
Glad I revisited this thread, I'd missed John's post on Shibano! Anyway, even though this position has many moves played, the LZ opening gospel can still provide guidance, as there a class 2 move available with that lonely 3-4 stone in a corner. The only caveat to the gospel that needs some judgement is "is there an urgent local situation?", it's easier to say no than yes, so then it recommends: The big low shimari is Elf's top choice, and the high is good too, and better than the human's move. Proselytize! Gospel: 2, Humans: 0, N/A: 1 |
|
| Author: | iopq [ Sat Jul 25, 2020 9:50 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
I thought it was black to move and I thought the correct thing is to play in the bottom right. I was right, fat KataGo wanted to 3-3 the bottom right with the win rate of 97% But since this is white to move, fat KataGo (40b 384c) likes these two moves Attachment:
File comment: two candidate moves Screenshot from 2020-07-26 00-36-00.png [ 921.02 KiB | Viewed 13449 times ] the first variation could go something like this Attachment:
File comment: first variation Screenshot from 2020-07-26 00-43-31.png [ 856.43 KiB | Viewed 13449 times ] the second variation could go something like this Attachment:
File comment: second variation Screenshot from 2020-07-26 00-49-12.png [ 852.98 KiB | Viewed 13449 times ] The percentages are roughly the same, the one it liked initially doesn't look as good if you explore the "main line" but still roughly the same |
|
| Author: | dhu163 [ Fri Jun 04, 2021 9:55 am ] |
| Post subject: | Re: Opening problems for AI: Problem 3 |
| Page 1 of 1 | All times are UTC - 8 hours [ DST ] |
| Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |
|
in the game.
and 
are from the actual game. Next, Black played at a, but Elf prefers to approach the top left corner, not allowing White to make an enclosure plus the extension from it.