Life In 19x19 http://www.lifein19x19.com/ |
|
Opening study with KataGo http://www.lifein19x19.com/viewtopic.php?f=48&t=17837 |
Page 2 of 6 |
Author: | lightvector [ Mon Nov 23, 2020 2:18 am ] |
Post subject: | Re: Opening study with KataGo |
Quote: Recalling KataGo's advice on position 3 of game 6, I chose to play c here. But KataGo says that is a bad move, and I should play a! I really don't see how a is different from the move I was admonished for playing in the previous game. I would be super happy if anyone has thoughts that can clarify this. So I guess, if we are to overgeneralize a little without checking other positions, then the TLDR for this is: the 3-4 point one-point-jump enclosure makes it the most important to reduce or expand on the side it faces. The 4-4 knight's move enclosure makes it the least important. The other enclosures are in-between. So in the first example, playing on the side was correct because white's enclosure was a 3-4 one-point-jump enclosure while black's was only a small knight's. In the second example, playing on the side is bad because white's enclosure is a 4-4 knight enclosure and limiting black is better because now black is the one with the 3-4 one-point-jump enclosure making the bottom side the most important. |
Author: | hakuseki [ Mon Nov 23, 2020 7:50 pm ] |
Post subject: | Re: Opening study with KataGo |
Hey, thanks a lot for all the comments and analysis! So, I think my tentative conclusion is that I need to compare the quality of the black and white frameworks to prioritize my own development vs. stifling the opponent's development. |
Author: | hakuseki [ Mon Nov 23, 2020 8:14 pm ] |
Post subject: | Re: Opening study with KataGo |
RobertJasiek wrote: "KataGo doesn't like this joseki" What is KataGo's difference of percentages at the game start and at this position from Black's view? And does it see the ko at this moment? Assuming a 6.5 komi and Japanese rules, white begins the game at 51.3%. After white's contact play, white is at 51.9%. When black plays the atari to enclose white, white's win rate jumps to 68.0%. With black's solid connection, white climbs to 75%. By the end of the sequence, white has dropped slightly to 74.4%. This is with the latest 30-block network on my machine with an imprecise number of evaluations. I don't usually talk about win rate in my analysis, though. I think score estimation is more intuitive, and also probably more stable across different networks and evaluation counts. I think it's a bit hard to say what KataGo thinks about the ko, exactly. It seems pretty sure that white will end up owning the corner, based on the expected territory view (I'm not sure how to pull actual numbers for this out of KaTrain, though, as it only shows me a visualization). I think that makes sense; starting the ko is more of an endgame move, so we will not see it happen for a long time. Maybe there's a 50% chance that black will start the ko (the other possibility being that white settles the corner first). Then white captures first, so I'd guess white has a >50% chance of winning the ko conditional on the ko starting. Overall white should have a >75% chance of living in the corner. |
Author: | hakuseki [ Mon Nov 23, 2020 9:20 pm ] |
Post subject: | Game 7, part 2 |
Game 7, position 2 Since I played the side extension, black was also allowed to establish a framework, as effectively predicted by several observant comments. I played ![]() ![]() It seems that my two-space jump was incorrect on two counts. First, the timing was wrong, and white should first push once (why not twice? I believe the first push is sente, as it threatens to hane the two stones, which is rather severe, but after ![]() ![]() ![]() At first I wondered if there's some kind of shape preference for a small or large knight's move over a two-space jump. But considering moves like a-d, with no clear shape relationship to any other stone on the board, I doubt that's the issue. Instead, I think it's just that next stone should be near the "middle" of white's framework, and N10 is not in the middle, being closer to the lower side. Game 7, position 3 My move was ![]() (game 7 to be continued) |
Author: | RobertJasiek [ Mon Nov 23, 2020 11:28 pm ] |
Post subject: | Re: Opening study with KataGo |
hakuseki wrote: Assuming a 6.5 komi and Japanese rules, white begins the game at 51.3%. [...] By the end of the sequence, white has [...] 74.4%. Thanks. I just wanted to know that White has a significantly larger percentage than before in KataGo's opinion. Quote: I think it's a bit hard to say what KataGo thinks about the ko, exactly. [...] Maybe there's a 50% chance that black will start the ko As long as we do not know if KataGo sees and evaluates the ko correctly, we cannot know if its percentages or score estimates are meaningful from a human view. The ko deserves an endgame-like evaluation of the local count (for the sake of simplicity, I exclude Black's outer territory), 1 white move away from almost -9 or 2 black moves away from ca. 24. The local move value is ca. (24 - (-9)) / 3 = 11. Hence, the local count of the corner territory is ca. -9 + 1 * 11 = 24 - 2 * 11 = 2 (favouring ---Black---!). Of course, the influences after either outcome play a great role, too. Black losing the ko loses much influence on the right side. |
Author: | hakuseki [ Tue Nov 24, 2020 1:14 am ] |
Post subject: | Game 7, part 3 |
Game 7, position 4 Actually, white has several good options in this position, such as a-e, but another excellent option is to attack the marked stones. What's the best move to do so? There are many possible continuations from here. Here are a couple examples: I am including these for illustrative purposes. However, there are many variations and I wouldn't expect to actually read them all out. Instead, I hope I would find this move by thinking in terms of shape. Black's marked stones are an elephant jump shape, which is often considered worth cutting. Furthermore, white's marked stones form a shape that I'd like to comment on. I have started to think of this shape as a "knight's triangle." Although this is the first time I have mentioned it here, I've already seen it appear quite a few times in various game reviews with KataGo. I believe it is a good shape. I still haven't ingrained the habit of looking for this shape, but I think it may be an important concept. |
Author: | Bill Spight [ Tue Nov 24, 2020 1:49 am ] |
Post subject: | Re: Game 7, part 3 |
hakuseki wrote: Black's marked stones are an elephant jump shape, which is often considered worth cutting. Furthermore, white's marked stones form a shape that I'd like to comment on. I have started to think of this shape as a "knight's triangle." Although this is the first time I have mentioned it here, I've already seen it appear quite a few times in various game reviews with KataGo. I believe it is a good shape. I still haven't ingrained the habit of looking for this shape, but I think it may be an important concept. I have been calling it the dogleg keima. ![]() |
Author: | hakuseki [ Tue Nov 24, 2020 4:50 pm ] |
Post subject: | Game 8, part 1 |
I played black in this game, and lost about 11 points in mistakes by move 50. Game 8, position 1 Here's the game start. Both players have approached each others' 3-4 points. I thought about the Leela Zero Opening Gospel -- but my recollection was imperfect -- and I played d. After reviewing the Gospel (which doesn't talk about the 5-4 approach explicitly), I think a better generalization would be that a follow-up to an approach is higher priority than a response to an approach, as KataGo seems to also believe here. The preferred moves are a or b, with e not far behind. I'm not sure why the keima is better than the kosumi in the upper right corner. Game 8, position 2 Narrowly considering the joseki in the top left, the side hoshi is the typical follow-up. However, I've gotten the impression that a wider extension is pretty common if it would be a multi-purpose move. For example, if the upper right corner had a black 4-4 stone, I've seen AIs finish this joseki with N17. So I figured a pincer would be correct here. Also, locally speaking, it seems to me that AIs don't usually play a pincer immediately when approached (unless this would block the formation of a large framework) -- instead favoring something like a kosumi or knight's move extension -- but if the opponent tenukis after this, a pincer is usually the best local follow-up. Specifically regarding the upper-right corner, ![]() But more recently I've noticed that if white tenukis, KataGo seems to prefer c as a follow-up for black. As for white responding locally, a and b seem to be equally good options. |
Author: | hakuseki [ Wed Nov 25, 2020 4:31 am ] |
Post subject: | Game 8, part 2 |
Game 8, position 3 White played the slide, and I unfortunately played c, which in this case is significantly worse. KataGo recommends a, but b is nearly as good, and to me seems a bit less adventurous. I'll plan to use either a or b if I reach this position in the future. Game 8, position 4 White approached my 3-3 stone, and we played the following sequence. What's next? A long time ago I liked a. But KataGo says b is better, so recently I've been playing b. Game 8, position 5 However, even though I've previously learned to play ![]() ![]() ![]() I think this is a pretty cool move. If white wants to block black's connection, it's obvious that a won't work. White's options are b and c. This position doesn't appear in Josekipedia, so let's look at the main variations here. Variation 1 ![]() ![]() There are actually many continuations to choose from, but this line for black seems to limit white's options the most. Variation 2 Variation 3 If white blocks at ![]() |
Author: | hakuseki [ Thu Nov 26, 2020 5:23 pm ] |
Post subject: | Game 9, part 1 |
This is a game I played on KGS recently. It was a bit dismaying to see that my mistake total in this game was much higher than in my practice games in KaTrain, even though I played significantly more slowly here. I guess playing AI is a bit different than playing human opponents. I played white in this game, and lost about 22 points to mistakes by move 50. Game 9, position 1 My opponent played the low Chinese opening. While I've seen this opening many times before, it has never come up in my opening practice, and I am not very familiar with AI opinions on this opening. No (white) mistakes so far. Game 9, position 2 A common pincer joseki. Yet not so common in my studies, because KataGo does not pincer very often. Where should white play -- a, b, c, or d? A long time ago, I thought that a was "the" move. Then I learned that b might be okay too. Later I learned how to respond to c, but never played it myself, as I thought it was a trick move. I've never seen d. It seems I had it backwards. KataGo recommands d and c about equally, with b being worse, and a being even worse. I guess d is already well known (just not to me), since it appears in Josekipedia. (Game 9 to be continued) |
Author: | hakuseki [ Thu Dec 03, 2020 11:14 am ] |
Post subject: | Game 10, part 1 |
I have a more recent game to discuss, so I'm setting game 9 aside for now. This is another online game. I was playing black. Game 10, position 1 In the past I've usually preferred the stability of an early enclosure, but I'm getting the sense that an approach like ![]() Game 10, position 2 ![]() ![]() ![]() ![]() ![]() Game 10, position 3 I recall that, without reading any actual sequences, I simply visualized the following result and concluded that it was bad, so I simply abandoned my splitting move and played a instead, thus throwing away about 2.5 points. This sequence should have been very easy for me to read, and the result seems happy for black; white has no territory and two cutting points. Game 10, position 4 ![]() ![]() ![]() a-d are some options given in policy order (which, it occurs to me, could be a nice convention for enumerating alternative moves). I played b, thinking about the knight's triangle shape. But based on the policy ordering I guess a might be the more intuitive "shape" move. Actually a-c are about equal in value. However, an even stronger option would be to cut and fight at d. |
Author: | hakuseki [ Wed Dec 09, 2020 9:42 pm ] |
Post subject: | Game 11 |
Diagram 0 I decided to aim for a high rank on BadukPop, so I'm reviewing my 9x9 games with KataGo. By the way, I played white in this game, and made mistakes totaling about 17 points. My opponent's mistakes totaled about 16 points, so I lost. Diagram 1 - The opening a-e are some possibilities for white's next move. This is in policy order (so, e.g. a is not necessarily the best move, but it is the most intuitive according to KataGo). Diagram 2 - First mistake This was my move in the game, but a mistake of approximately 1 point. I think creating a separate group here is a poor choice. Actually, move a from diagram 1 would have been the correct answer. I did consider a, but was worried about black cutting. So my actual mistake was not reading that black's cut would fail. This is a sequence I should be able to read. Actually, it is not best play for either player, but it convinces me that black's cut fails. Caveats:
Actually, this cut did end up getting played later in a somewhat altered situation, which resulted in more mistakes during the game. But I believe those are redundant from a review perspective. Diagram 3 - Endgame problem 1 Endgame move suggestions for white are a-f, in policy order. Diagram 4 - Second mistake I played ![]() Although there are several possibilities to check, perhaps the most obvious idea for black is this wedge at ![]() ![]() ![]() Diagram 5 - Endgame Problem 2 Diagram 6 - Third mistake This is what I played in the game. It achieves two goals -- it connects the white stones, so that white will not need to play a future move at a, and it also reduces black's liberties, so that b may later be played in sente. However, this is not the best sequence for white. Actually, the solution to diagram 5 was c, a brilliant tesuji. Diagram 7 - Shining tesuji If white is allowed to follow up at b, the presence of ![]() Diagram 8 - Endgame problem 3 Move suggestions a-c are in policy order. Diagram 9 - Fourth Mistake This is the game sequence. At this point I realized that connecting at a would allow black to invade the corner and create a seki. Evidently, I know this corner shape well enough to recognize the weakness once it appears, but not quite well enough to prevent it in advance. Naturally, the correct answer in diagram 8 would have been a. |
Author: | hakuseki [ Thu Dec 17, 2020 4:18 am ] |
Post subject: | Game 12 |
This is a 9x9 game played on BadukPop. I played black. Game 12, position 1 In this position my move was e, but I think that was an overplay. KataGo's policy suggests moves like a-d. Interestingly, the score evaluation for these moves fluctuated a lot as I let it run for about 100k playouts, but generally a seems to be the best, and at the end of the evaluation it was favored by >1 point over any other move. Black sacrifices ![]() ![]() ![]() ![]() This isn't the first time I've seen KataGo suggest something like this, yet I virtually never play this way. I did consider playing ![]() Since my hypothesis is that ![]() ![]() ![]() ![]() ![]() Here we see that ![]() ![]() ![]() I guess this seems to verify the purpose of the suggested move a in position 1. Maybe I can try to keep an eye out for this kind of move. |
Author: | hakuseki [ Sun Jan 03, 2021 11:47 pm ] |
Post subject: | Game 13 |
A new 19x19 opening after a couple weeks of 9x9. I'm black. I lost about 40 points by move 80. I don't see the two-space high approach much. ![]() ![]() When white pincered at ![]() ![]() I answered ![]() ![]() I think this is makes sense to me in retrospect. The 2-space approach is unusual, so it's reasonable to assume it's worse than the typical 1-space approach. But ![]() This is likely to lead to the large avalance joseki. |
Author: | hakuseki [ Mon Jan 04, 2021 12:07 am ] |
Post subject: | Game 14 |
I played white in this 4-stone handicap game. By move 80 I lost about 25.5 points. Game 14, position 1 The game started with this pincer joseki in the upper right. This is pretty common, yet I still hadn't learned the continuation. Continuation In the game I played elsewhere instead of playing this sequence. This is especially sad because I recently watched a video where this was mentioned, but I wasn't really paying attention. I think I'll remember it next time, though. Missing this sequence cost me about 1.4 points in KataGo's estimation; it also accounts for about 3.3 points of the 25.5 point total mentioned above, as I missed multiple opportunities to start the sequence. |
Author: | hakuseki [ Mon Jan 04, 2021 1:54 am ] |
Post subject: | Game 15 |
I played another 4H game as white. This time I lost 81.1 points (oops!) by move 80. Game 15, position 1 One nice thing about 4H games is that KataGo plays a lot of pincers as black. I think this will provide valuable practice, as KataGo rarely pincers in even games, but my human opponents pincer often. Here's a pincer joseki that also appeared in game 9. In that game, KataGo suggested that a and b were nearly equally-good continuations. Today I played a. However, in this 4H game position, KataGo really disliked a. I have yet to see a position where b is considered bad, so I will make it my new default choice. Game 15, position 2 This is the game continuation. After ![]() ![]() Game 15, position 3 In the game I played f, which was not so good. KaTrain on my machine ranks the moves in order e, a, c, b, g, f, h, d. I think I should have played either a or e, but didn't. So let's look at how I read these two moves. Variation a I didn't play ![]() ![]() Fortunately, white can do much better than this, as both of these white responses are mistakes. Starting with ![]() White can just keep extending, and now black's corner becomes very weak. Or, even better: White can enclose the top. Black's stones are not dead yet, but black will now be thinking about the best way to sacrifice them, rather than chase white into the center. (hopefully to be continued) |
Author: | hakuseki [ Tue Jan 05, 2021 10:03 pm ] |
Post subject: | Game 16 |
I played black in this game, and made 24.5 points of mistakes by move 80. My AI opponent in this game played by stone-scoring rules. However, I'm not currently currently trying to learn strategies under group tax (if I do so, I will start a separate thread for that); I'm still using Japanese rules for analysis, just against a different style of opponent play. Game 16, position 1 This position was discussed in a variation on game 5. At this point, white should play the ladder breaker at a... Game 16, position 2 Instead, white ataris first, and then plays the ladder breaker. This is a mistake, because the ![]() ![]() ![]() Game 16, position 3 In this position I played a, which is a perfectly good move, and in fact KataGo's preferred policy move. However, I want to note that b is also perfectly good, and seems to remove more weakness than a. Game 16, position 4 My move here was c, but this is losing one point. The correct move is a: ![]() ![]() ![]() Game 16, position 5 This seems to be a modern joseki, although ![]() I wasn't familiar with the joseki and played c, but this mistake cost me 1.5 points. The correct answer is a: Game 16, position 6 Finally, here's a whole-board puzzle. Black would like to reduce the upper left. After 100k visits on my machine, KataGo chose a move with a policy value below 1%, yet preferred by more than a full point over any other move. |
Author: | RobertJasiek [ Tue Jan 05, 2021 10:27 pm ] |
Post subject: | Re: Game 16 |
hakuseki wrote: This is an old joseki but not known / taught by everybody. |
Author: | hakuseki [ Wed Jan 06, 2021 9:28 pm ] |
Post subject: | Game 17 |
Another game as white against the stone-scoring AI. I'll try focusing more on 4-4 against this AI. As before, I'm still doing the review without group tax. In this game I lost 32.2 points by move 80. Game 17, position 1 ![]() Game 17, position 2 Since ![]() Game 17, position 3 The pincer at a is really good, but more importantly extending at d, which was my move, is really bad. I guess a typical joseki move might make a weak stone strong, while this move merely makes a very weak stone into a somewhat weak group. Locally, c would be about 0.6 points better, while a would be a further 0.2 point improvement. Game 17, position 4 Here I played c. I was probably thinking about getting a few points of territory and simultaneously robbing black of a few points of territory. At the same time I would be helping my isolated stone at C8. Aside from the fact that a is quite big, if I want to help the C8 stone I should actually play at b, as this simultaneously attacks black's weak group. This would be an improvement of about 0.6 points. Game 17, position 5 I had been thinking about the hane at d, and I regretted missing my chance once black played ![]() If I were really so sad about missing d, I could just play d now instead of f, and it would be an improvement of about 0.4 points. Or, if my plan is to prevent white from enclosing territory on the bottom, the knight's move slide at e would be a further 0.35 points better. Or, if I want to damage the black stones, splitting at a would be a further 0.9 points better. Or, if I want to learn a cool joseki followup, b would be a further 0.5 point improvement, threatening to either connect to the nearby white stones or jump into the corner. Or, if I want to learn another cool joseki followup, the clamp at c is another 0.15 point improvement, mostly likely allowing white to build some center influence in sente. All told, white can do about 2.3 points better than f. |
Author: | hakuseki [ Thu Jan 07, 2021 12:04 am ] |
Post subject: | Game 18 |
This time I played black and lost about 34.8 points by move 80. Game 18, position 1 Time to learn about the large knight's move extension. In this game, I tried playing in much the same way I would play against the small knight's move: This is the game result. Although neither black nor white should play this way, black came out worse, losing 1.4 points overall. It seems that white's marked stone is more efficient than in the standard knight's move joseki. In fact, white's preferences for where to put the marked stone seem to be a, b, c, d, its current location, e, and f, in that order. So what should be done differently? For white, this is an even better way to punish ![]() So black should simply jump into the corner, resulting in something like this sequence. (Game 18 to be continued?) |
Page 2 of 6 | All times are UTC - 8 hours [ DST ] |
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |