Life In 19x19 http://www.lifein19x19.com/ |
|
This 'n' that http://www.lifein19x19.com/viewtopic.php?f=12&t=12327 |
Page 27 of 53 |
Author: | Kirby [ Mon Nov 26, 2018 10:55 pm ] |
Post subject: | Re: This 'n' that |
Thank you, Bill, for the pictures. It's nice to put some names to faces. |
Author: | Bill Spight [ Thu Feb 14, 2019 10:40 am ] |
Post subject: | Re: This 'n' that |
Dear Friends and Gentle Hearts, A little Valentine's present. https://www.youtube.com/watch?v=WWLPwEF ... gs=pl%2Cwn |
Author: | Bill Spight [ Sun Mar 31, 2019 2:56 pm ] |
Post subject: | Re: This 'n' that |
Many thanks to Daniel Hu for his thoughts about AlphaGo and strategy. viewtopic.php?f=10&t=16527 I think that we humans will come up new and useful ideas and concepts about go, based upon the play and opinions of bots, even if they cannot explain their plays and preferences. Hats off to Daniel for making steps in that direction. I doubt if Elf (version 1, anyway) is as strong as AlphaGo, but the Elf team has given us a lot of material to study with its commentary on most of the GoGoD games. The commentary does not include handicap games, and its view of even games without komi is, perforce, flawed, as it was trained on 7.5 komi. I have also found a ladder mistake, BTW. One remark in the paper by the Elf team has made a slight difference in my view of the frequent tenukis of Zero bots. It was that Elf learned go backwards, in a sense, because it learned to predict the winner of the game, and early in its training it was, OC, better at that near the end of the game than early in the game. (This does not prevent it from making wacko endgame plays, OC. ) I filled in the blanks a little bit in that explanation, BTW. The endgame, as we humans know, is the phase of the game in which, except for ko positions, the board is divided into independent or nearly independent regions. In each of these regions one play, or maybe a few plays are made, and then a player tenukis, because plays in that region have become worth less, or, as we say, the local temperature has dropped. It is quite possible, then, that zero bots, who have not been trained on human games, view the go board is a number of quasi-independent regions, even in the opening. (I do so, myself, BTW. ) Perhaps they do so even more than humans. Quien sabe? If so, that would explain their greater propensity to play tenuki than humans have. Are humans wrong not to tenuki so much? Maybe, maybe not. To understand one reason why not we need to address the question of the margin of error of the bots' evaluations. (Nobody knows what the margin of error is, OC, because it has not been studied directly. IMHO, the margin of error of the winrate estimates of the AlphaGo teaching tool is at least 2%, and that of the winrate estimates of the Elf commentary is around 3%.) That is, if you are a human and Elf or Leela Zero says that you have made a 2% mistake, you may well have made the right play, and even if your move was an "inaccuracy", it may not cost any points with perfect play afterwards, and even if it does cost a point or two, it may not cost the game. That is especially so in the opening. OC, the bot as a player has to trust its evaluations, in accordance with its algorithms. So the bot may well tenuki (correctly) even when it does not actually need to do so, and a human in the same position would not be inclined to tenuki. ---- Here is a position when Takagawa, a player who understood the opening about as well as any human in the 20th century did, failed to tenuki, and Elf thinks that it was a 9% error. I'll go with Elf on this one. 9% is surely outside of its margin of error. You can probably guess Takagawa's error, since he did not have many opportunities to make one. (It was .) But where should he have played , and why, according to Elf? This was also the game where Elf missed the ladder. Maeda played as a ladder breaker. , OC, restored the ladder. Elf wants to play at 28. Edit: Let me add this. One Elf variation has at 30, restoring the ladder, but this one has White play instead of running out in the broken ladder. (This variations has only a small change in winrate.) |
Author: | John Fairbairn [ Mon Apr 01, 2019 2:37 am ] |
Post subject: | Re: This 'n' that |
Hi, Bill I have printed out Daniel's paper to read on my next train journey, but as you have endorsed it I shall try to get to it sooner! I have a few remarks below on what you said, which obviously are made in ignorance of what Daniel may have said. First, the trivial points. The move order shown for the Takagawa game (vs Maeda) is wrong, but of more substance is the fact that komi was only 4.5. Given that, which you correctly (as confirmed by the Elf programmer) see as a flaw, plus an apparent ladder error by Elf, I'm wondering how much we can rely on Elf being right here and Takagawa wrong. Now the meatier points. You use the word thickness about Takagawa's play. It has been my impression that he himself tended to use that term far less than other players (and conversely used terms such a balance much more than other players). I put it to you that it's not real thickness (it's gaisei) and so the "rules" about local plays around thickness (such as extensions) actually tended to be either ignored by Takagawa or were subsumed in his views about balance. I think that has obvious implications for how we view AI moves in the whole nexus spanning thickness, influence, walls, thinness, etc. Even in the traditional human thinking on the topic, there is room for deeper consideration (e.g. the fact that thinness is not actually the antonym of thickness), but in any case the heuristics for playing around influence differ from playing around true thickness. I suggest bots may be doing that more reliably than humans. Further (to take up your point about regions), one of the main tenets of New Fuseki theory was equilibrium and averaging based on symmetries (which, too, has a bearing on balance). Theorists then therefore made much of dividing the board diagonally rather than orthogonally, yet that idea seems to have been largely forgotten. Players today (even pros) seem focused still on viewing the board as four square quadrants, which I think is partly to do with obsession about josekis. But if you view the board as four equilateral triangles you get a very different view - perhaps one more like the one the bots are "seeing" but which may still be within the purview of humans? |
Author: | Aidoneus [ Mon Apr 01, 2019 6:15 am ] |
Post subject: | Re: This 'n' that |
Hi, John Would you care to expound a bit on how viewing the board split diagonally might lead to some different assessments? Perhaps you could point to some theoretical works translated into English? |
Author: | jlt [ Mon Apr 01, 2019 6:35 am ] |
Post subject: | Re: This 'n' that |
John Fairbairn wrote: But if you view the board as four equilateral triangles I usually don't view the board this way. |
Author: | John Fairbairn [ Mon Apr 01, 2019 10:03 am ] |
Post subject: | Re: This 'n' that |
Sorry, I was slipshod and used equilateral for isosceles. Like numbers, geometry falls into the class of things I have to stop and think about. The advantage is that I can often get away with bigger slices of pie. I have written a lot about New Fuseki theory in Old Fuseki vs New Fuseki. |
Author: | Bill Spight [ Mon Apr 01, 2019 10:51 am ] |
Post subject: | Re: This 'n' that |
Dear John, I think that we are going to have to rethink our ideas of thickness and influence because of go AI. And other concepts, as well. As for dividing the board diagonally, didn't you say that Sonoda sometimes does that? |
Author: | Aidoneus [ Mon Apr 01, 2019 12:31 pm ] |
Post subject: | Re: This 'n' that |
John Fairbairn wrote: I have written a lot about New Fuseki theory in Old Fuseki vs New Fuseki. $168.54 for a used copy from Amazon is just a bit too steep for me. Edit: I now have found the thread on your book. Too bad it seems to be out of print. |
Author: | Uberdude [ Mon Apr 01, 2019 2:27 pm ] |
Post subject: | Re: This 'n' that |
https://www.schaakengo.nl/goshop-keima/ ... -fairbairn €10 plus postage |
Author: | Aidoneus [ Mon Apr 01, 2019 2:42 pm ] |
Post subject: | Re: This 'n' that |
Uberdude wrote: https://www.schaakengo.nl/goshop-keima/books/books-by-publisher/slate-and-shell/ss59-old-fuseki-vs-new-fuseki-fairbairn €10 plus postage Thanks! Edit: I just discovered that the book is available from SmartGo, so I bought the electronic version. Which I can begin reading right now. |
Author: | Bill Spight [ Mon Apr 01, 2019 4:00 pm ] |
Post subject: | Re: This 'n' that |
Very strange. In note #534 I wrote that White played , and when I go to edit it, that shows up, but in the note as it appears, it says that green played . Does the same thing happen here? Edit: Yes, it does. What about Black? Edit: Black shows up as Purple. Later edit: Corrected now. Thanks. |
Author: | Bill Spight [ Sun Apr 07, 2019 9:15 pm ] |
Post subject: | Re: This 'n' that |
Direction of Play? Basic diagram: Just about every opening textbook has a diagram like this. The arrows indicate the direction in which each enclosure naturally develops. If Black plays on the left side first, she can play on C-10 and then, if White plays on C-12, Black can jump to E-10, forming an ideal box framework. Below are positions from actual pro games with enclosure facing enclosure on one side. The textbook play appears to be the extension to the 10-3 point. Is that where the pros actually played? And is there a better play? Since the title has a question mark, you might think that there is a better play. Anyway, food for thought. I'll reserve the next four posts for comments based upon the published Elf analysis. Note that three of the games have no komi. In a no komi game the best practical play for White may be to make an overplay, an aggressive play, or to complicate the game. Conversely, Black may play safe, sound, and simple. Elf, OC, assumes a 7.5 komi. Edit: It may come as no surprise that all the pros played the 10-3 extension, and that Elf preferred a different play in each case, to varying degrees. White to play. 4.5 komi. White to play. No komi. White to play. No komi. Black to play. No komi. |
Author: | Bill Spight [ Sun Apr 07, 2019 10:34 pm ] |
Post subject: | Re: This 'n' that |
White to play. 4.5 komi. Takagawa (of course, to those familiar with his play) took the C-10 point. Here are the next 10 plays. extended on the top side, preventing a double wing formation by White. Play continued in a fairly standard way for the time. ---- Elf evaluated as losing 4%. But Elf's margin of error is around 4%, I think. It's high because Elf gives large error estimates, not because its winrate estimates are bad. So I think we can say that Elf considers to be questionable. To help us understand Elf's evaluation, let's compare Elf's main line after to its main line after . Elf thinks that should attach to the top of (marked)! After let's evaluate the left side. Maybe White has gained about 1 pt. of territory, 2 pts. of actual territory for White vs. 1 pt. of central influence for Black. Compare that with the next diagram. With this exchange, White's moyo has gained around 8 pts. more than Black's has. OC, moyo is not territory, so maybe White has gained about 3 pts. more than Black, a couple of more points than in Elf's main line diagram. White plays his preferred joseki in the bottom right. (Other bots might play the kosumi or keima.) Then Black plays the Go Seigen/AlphaGo shoulder blow, rather than an extension on the top side. (Well if Elf does not like the extension on the left side, why would it like one on the top side? ) surprised me. Since reduces the White moyo, why extend it? But AlphaGo, Elf, and other bots like the center. And this move nullifies Black's advantage there. (In particular it prevents a two step hane in the center. It also prevents Black E-13, threatening to connect either way.) and strengthen Black on the top side while still reducing the White moyo. Note that is an approach, not a wedge or pincer. Now let's take a look at Elf's main line after . White plays his preferred joseki in the bottom right corner. Then Black plays an ogeima enclosure in the top right, distantly approaching the bottom right stones. "a" is a rival with approximately the same win rate estimate, but "a" only gets 26k playouts to 51k playouts for the enclosure. approaches the top right corner. "a" and "b" have the same win rate estimates, but "a" gets only 17k playouts and "b" gets 5k playouts, while gets 22k playouts. Surely the left side extensions are playable at that point. is a close pincer. "c" is a rival, but gets only 3k playouts versus 17k. finally makes the extension on the left side. The pincers "a" and "b" are rivals, with the same win rate estimates, but again it is a playout contest. "b" gets 3760 playouts and "a" gets 5230, while gets 5313, a very close call. plays the shoulder hit and taps it underneath to contest the top side. ---- What seems to be going on is that Elf considers the extension for White to be premature at move 8, and perhaps prefers the attachment against the top left enclosure for Black. Perhaps it is an example of the dictum, corners before sides. Corners may remain urgent longer that we had thought. And the top attachment against the White extension was an eye opener for me. It indicates that the extension may not be as large as the textbooks suggest. Edit: Please note that the board in the actual game is rather more developed than in either of Elf's main lines. |
Author: | Bill Spight [ Sun Apr 07, 2019 10:35 pm ] |
Post subject: | Re: This 'n' that |
White to play. No komi. Fujisawa Hideyuki (AKA Shuko) also played the 10-3 extension. The subsequent play has some notable features. combined a 5 pt. extension from the enclosure with a 3 pt. extension from the Black stone in the top left corner. completed a double wing formation from the bottom left corner. When I was learning go, a double wing was supposed to be ideal, but pros, both at the time and historically, had not always rushed to prevent one. W"a" is a common response to the pincer, but Fujisawa chose to play a Kitani joseki, settling White into the corner. Then was a very high reduction of White's double wing, an unusual choice. Black may have been counting on the lack of komi to keep White from playing strongly against the reduction in exchange for a bit more territory. ---- Elf considers to be an 8% mistake, so it very probably is a mistake. As above, let's look at Elf's main line after and compare it with its main line after . Black plays the taisha in the top left corner. White takes territory in exchange for outside strength. is a tesuji noted not long ago by John Fairbairn in human play in a different position. After Black retains sente and approaches the bottom right corner with . This approach to the high enclosure has a long pedigree. Now let's look at Elf's main line after . As we may have anticipated, Elf considers the top left corner to be the hot spot of the board. is the now standard keima. Black turns to the right side, but starts with the outside attachment to White's bottom right enclosure. After that kikashi Black extends to the 10-4 point. Again, "a" has the same win rate estimate, but loses the playouts, 5k to 22k. To me that makes sense, because the high extension looks good with the rest of the board, and because of the Black stones in the bottom right, Black is not concerned about a White play at "b". invades the bottom left corner. Black takes sente and returns to the the bottom right with the turn, . is a tesuji aiming to play at "a" or "b". As is often the case, Elf ends its variation in media res, leaving us to wonder how Black will reply. My main takeaway is similar to the last example, The corner is hotter than the side, even one with facing enclosures. This corner, with the 3-5 Black stone on the side closer to the Black enclosure instead of the lone 4-4 stone, seems to be hotter than the one in the Takagawa game, perhaps because a Black wall would work better with the enclosure, so that Elf considers the side extension to be a worse mistake. BTW, Elf considers in the actual game to be a 9-10% mistake, even worse than , returning the favor with interest. Both players, despite being among the world's best at the time, seemed to share an illusion about the value of those extensions, by comparison with a play in the top left corner. |
Author: | Bill Spight [ Sun Apr 07, 2019 10:37 pm ] |
Post subject: | Re: This 'n' that |
{Under construction} White to play. No komi. Cho Chikun also played the 10-3 extension. Then Kono played the usual underneath attachment in the top left corner. The kosumi was standard when Black also had a strong, low stone to the right ( ), to provide a high-low balance. was the standard blocking extension on the right side. Cho approached the bottom left corner and then made a double wing formation with was a probe in the corner, and after pulled back, Kono played the invasion at . {More analysis to follow.} |
Author: | Bill Spight [ Sun Apr 07, 2019 10:41 pm ] |
Post subject: | Re: This 'n' that |
{Under construction} Black to play. No komi. Shirae also played the 10-3 extension with . Sugiuchi (still active at age 92, see the Sugiuchi watch thread, viewtopic.php?p=241376#p241376 ) replied with , the standard checking extension on the right, only a two space extension because of the weakness of the ogeima enclosure. Black then pincered on the left side, and the battle continued there for several moves. {More analysis to follow.} |
Author: | EdLee [ Mon Apr 08, 2019 3:09 am ] |
Post subject: | |
Hi Bill, Curious about a low-end bot's evals. Takagawa-Tainaka: |
Author: | Bill Spight [ Mon Apr 08, 2019 4:45 am ] |
Post subject: | Re: |
EdLee wrote: Curious about a low-end bot's evals. Interesting. Since Deep Leela was trained on human play, I thought that it would agree with a human at least once. Quote: Cho-Kono (small typo in #534's date: '1907' ) My bad. OTOH, the spell checker changed Cho's name to Cho Chicken. |
Author: | John Fairbairn [ Mon Apr 08, 2019 11:47 am ] |
Post subject: | Re: This 'n' that |
Bill Just as background for your splendid efforts. I have been trying to come to terms with what I thought was something odd about facing shimaris in pro play. Like you, I think, I was brought up in go during the Takagawa era. What I learnt then was that the main direction of play from the ikken tobi shimari was along the side that makes an L shape. The reason was that this leads to a box shape rather than a tray shape. I have been increasingly puzzled lately (even before AI times) as I have noticed more and more pros favouring the other ("open") side. I cannot recall seeing anything about this in fuseki books. This is so extreme that (from memory) of over 200 games that I have with facing ikken tobi shimaris, in only two, I think, was the centre mid-point played. Where one kogeima shimari faces an ikken tobi shimari, the centre mid-point play was rather more common, but still close to disappeared from pro play around 2010, or even earlier. But - big BUT - where a kogeima shimari faces another kogeima shimari, the centre mid-point is a very common play even today. That has long baffled me. But in the light of what the bots have shown us, one vaguely possible thought has come to me: a contact play at the third-line root of the shimari seems powerful with a kogeima shimari, but less so against the ikken tobi shimari. Is that side therefore favoured for an extension from the kogeima simply to avoid that sort of contact play. I have seen nothing in print to support that, but I do recall that Yi Ch'ang-ho led a fashion for playing a very early contact play on the second line. Could it be that he was on the same lines as the bots but just didn't realise the third line was better than the second for the contact play? |
Page 27 of 53 | All times are UTC - 8 hours [ DST ] |
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |