| Life In 19x19 http://www.lifein19x19.com/ |
|
| Retraining the Neural Network in My Head - Part 2 http://www.lifein19x19.com/viewtopic.php?f=15&t=18144 |
Page 1 of 1 |
| Author: | thirdfogie [ Thu Apr 08, 2021 6:10 am ] |
| Post subject: | Retraining the Neural Network in My Head - Part 2 |
This post follows on from the "which side to play atari" post. It shows situations where White needs to save one or more stones in atari. He must choose to extend or capture. The White player (me) usually makes the wrong choice according to KataGo. In the above diagram, the circled stone is in atari. I played at a, but b is 8.9 points better according to Katago. I was worried about the potential (indirect?) ko if Black replies to b at a, but failed to notice that Black can gain a liberty in the capturing race after White a by playing b himself, which gives Black a better squeeze on the squared white stones. In the above diagram, the marked stones are in atari. I played a, but b is 0.8 points better according to Katago. That is a significant difference this late in the game. The reason is that White b forces Black c, which I had not noticed. If Black does not answer b at c, White makes further gains by pushing in at d. I often have trouble seeing what is what late in the game when the visual field is busy, as here. This is true whether playing online or over the board. Katago sees clearly and invites me to think. In the above diagram, the marked stone is in atari. I played a, which is 1.5 points better than b. However, KataGo thinks that White c is best here. (Even when you are right, you are still wrong.) In the above diagram, the marked stone is in atari. I played a, but b is 3.8 points better according to Katago. This shape often occurs on the second and third lines, and I have noticed that strong players usually capture the opponent's stone. In the above diagram, the marked stone is in atari. I played a, but b is 2.4 points better according to Katago. The situation is similar to that in the previous diagram. Review There is not enough material to draw any general conclusions. However, accurate reading always helps, and one should probably capture any cutting stone on the third line immediately if possible. |
|
| Author: | dfan [ Thu Apr 08, 2021 6:57 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
thirdfogie wrote: There is not enough material to draw any general conclusions. However, accurate reading always helps, and one should probably capture any cutting stone on the third line immediately if possible. In these sorts of atari situations, I believe that in general stronger (let's say 1d) players are much more likely to just capture a stone, and weaker (let's say 5k) players are more likely to extend instead of taking the stone off the board, which looks more efficient but often leads to the opponent getting forcing moves later or having some aji still around. I think that it is a good instinct for players of our level to first consider making the capture, unless there is clearly zero aji. |
|
| Author: | jlt [ Thu Apr 08, 2021 7:34 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
To decide whether to extend or capture, I try to imagine the followup. Consider the following diagram. If White plays at "a" then  is forcing.  plays elsewhere because any local move is gote, then White comes back later to play  with sente.If White plays at "b" then plays elsewhere, but then White has sente endgame moves. |
|
| Author: | dfan [ Thu Apr 08, 2021 10:21 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
jlt wrote: If White plays at "b" then plays elsewhere, but then White has sente endgame moves.I believe that for this reason Black usually forces with first, since White won't be able to make as much progress with a in the following diagram:Confirmation or disagreement from stronger players would be appreciated! |
|
| Author: | jlt [ Thu Apr 08, 2021 10:30 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
I am not sure that in your diagram is forcing. It's big, but how big compared to the rest of the board, I don't know.
|
|
| Author: | dfan [ Thu Apr 08, 2021 10:37 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Good point. I just asked KataGo now and it wants to first play the double sente sequence starting with E2, then come back and play my .
|
|
| Author: | thirdfogie [ Thu Apr 08, 2021 11:33 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
It was the prospect of a ko if white plays elsewhere after that made me extendinstead of capturing during the game. Extending was clearly a mistake in this game because White is so strong to the left of the potential ko that he has nothing to fear from Black starting it. In the general case, it feels burdensome to have to evaluate a potential ko before making the "extend or capture" decision, so I will adopt the "Ape stronger players" policy and hope to learn by experience if any exceptions do appear. |
|
| Author: | dhu163 [ Thu Apr 08, 2021 1:54 pm ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
next diagram - interesting. Difficult problem. next diagram - explained very well by others. Even if black were solid elsewhere on the board, I would be very reluctant to sacrifice the chance to get into black's upper right corner just to avoid a (medium size) ko that black has to spend 2 moves to start. The ko isn't that small, but black is so thin in the upper right. In this game, black's shape on the left seems to have some problems and some ko threats, or at least threats to set up ko threats (in order to profit), or to get endgame such as E2. |
|
| Author: | John Fairbairn [ Fri Apr 09, 2021 6:40 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
An interesting exercise. Quote: The White player (me) usually makes the wrong choice according to KataGo. Have you got any inkling as to why this bias may be so. E.g. did someone give you bum information in the past, or did you lack playing over enough pro games (i.e. you did not previously train your neural network on the basis of pro games)? Or perhaps there are other psychological factors, such a being more concerned with depriving the opponent of something rather than attending to your own concerns? (I happen to believe this last point is something that applies to ALL amateur play.) |
|
| Author: | thirdfogie [ Sat Apr 10, 2021 1:39 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Thanks to everyone who commented. John Fairbairn It feels like the mistakes spring from the last of your causes: giving too little priority to weaknesses in my own positions. I haven't played over any professional games on a real board since September 2013, when my wife was away for a month. Watching professional games online is clearly not the same thing. We are told that passive play is a sure road to defeat, and it may be that I take that advice too simplistically. Sometimes, "reculer pour mieux sauter" should be the motto. Or one could simply blame egotism. dhu163 Thanks for the detailed analyses. I am amazed that anyone could apply end-game mathematics to the game shown below, given that he must first identify all the candidate sequences. On the other hand, people have composed great music using six-part counterpoint and performed highly abstract mathematics in a way which makes me gasp for breath, so what do I know? The diagram shows KataGo's best moves following White's mistake. Model b40c256-s6485784576-d1573360039 was used with about 30,000 visits per move, though the number of visits seems to make little difference this late in the game. White has a lead of about 11 points, but KataGo still picks the mathematically best moves as far as I can tell. In the actual game, black played 1 and I replied with a grotesque self-atari (white at black 5). After that, I was lucky to win by 2 points. In the following game, my original comment on the situation was wrong. I should have analysed more with KataGo before posting. The continuation is what actually happened. Black's terrible mistake at 5 may be instructive for some. |
|
| Author: | John Fairbairn [ Sat Apr 10, 2021 3:09 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Quote: We are told that passive play is a sure road to defeat, and it may be that I take that advice too simplistically. Sometimes, "reculer pour mieux sauter" should be the motto. Or one could simply blame egotism. This interests me because it is a language thing rather than a maths thing. Without getting into Sapir-Whorf arguments I do believe there is a link between language and actions. The reason this particular post of yours interests me is that you say someone has been telling you that passive play is to be avoided. But who? I've racked my brains but I can't recall the word passive being used in oriental texts. Very occasionally there are words that tend in that direction (e.g. "negative"), but these are very occasional, and my gut feeling is that such comments are applied to beginners far, far more than to stronger players. But, in contrast, what I notice is ultra-common in pro talk are references to timing. To be specific, if we take the atari on the outside and the atari on the inside in your first example, what came over to me was an attempt to say one is better than the other in some static way. In your case this may be because you apparently perceive one as potentially aggressive or passive, but I could easily imagine others may say it's a matter of good and bad shape, for example. However, my impression from pro talk is that neither is better or worse than the other in such terms. Nor is it, perhaps strangely, a matter of reading. Rather, it is a matter of timing. We can infer this from the fact there are so many comments of the type "Black should first play A, then B". In other words, purely in shape terms B is not at all inferior - it's just inferior at this particular moment in terms of timing. Furthermore, this can be but does not have to be the same as gote no sente (or reculer pour mieux sauter). That is, timing issues are pervasive - they are not just honte moments. As some mild proof of that, I used Kombilo to check the position of two different-coloured ikken-tobis side by side with one side wedging between the opposing ikken. Kombilo came up with lots of examples in the 110,000-game GoGo database - and the inside and outside ataris as played by pros came up exactly even! In short, I think we vastly overate shape and underestimate timing in go. It may be time to revisit Juergen Mattern's Der richtige Zeitpunky? It's a very long time since I read that and I can't remember what's in it. Does anyone have a more recent memory? But, in general, what do you think of exploring the notion of timing more. |
|
| Author: | Uberdude [ Sat Apr 10, 2021 2:36 pm ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
John Fairbairn wrote: Quote: We are told that passive play is a sure road to defeat, and it may be that I take that advice too simplistically. Sometimes, "reculer pour mieux sauter" should be the motto. Or one could simply blame egotism. <snip> But, in contrast, what I notice is ultra-common in pro talk are references to timing. I think the relative abundance of discussions on timing in pro commentaries in pro games, does not necessarily mean it should be of the same abundance and importance in commentaries on kyu games. Kyus make gross mistakes (chief categories being massive underplays aka passive play, and also massive overplays / over-aggression) that pros left behind many years ago when still little kids. So of course pros aren't going to be talking about things like wasting a move to live with a group that was already alive, or how to kill a bulky 5, or not playing a 2 point move in the opening when there are 30 point moves up for grabs, because that's super easy and obvious for them. Timing is one of the more advanced topics which they can still get wrong and differ in opinions, so they talk about it more. This is not to discourage more talk of timing here, it's a fascinating topic and indeed one of my favourite books "Beyond forcing moves" could be subtitled "the importance of timing". Here's a thread about a critical timing issue in a common post-joseki joseki https://lifein19x19.com/viewtopic.php?f=17&t=5150. I find it interesting that thirdfogie says he was trying to avoid passive play as his justification for the extend here (or maybe in general and not this example?): I would characterise this mistake as being passive! To illustrate, with a simple tewari: white instead captures at 1, black descends at 2, white then answers at 3, kindly giving black the large endgame move at 2 in sente. 3 is, I think we can all agree, terribly passive, much better to tenuki and play elsewhere. But by extending at 3 first, this is the result you get. So the only reason to not capture for 1, is if you think black won't play descend for 2, but will play the atari for a ko and you don't like that. Quote: It was the prospect of a ko if white plays elsewhere after that made me extendinstead of capturing during the game. But if you are scared of ko don't play go! as the saying goes. So I would actually say the descend is more accurately described not as avoiding passive play, but avoiding a scary looking ko (which stronger players know isn't scary, because wimping out of that ko is admitting defeat without a fight). |
|
| Author: | dhu163 [ Sun Apr 11, 2021 5:29 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Quote: I am amazed that anyone could apply end-game mathematics to the game shown below, given that he must first identify all the candidate sequences. Its late enough in the game that I think CGT theorists could solve it perfectly with some time and maybe a computer. As for me: Well, the KG diagram means I got it wrong all over the place. I haven't put in enough study of the endgame myself and I'm learning here too. I normally only use approximations, which is insufficient for accuracy on the whole board. But there are also several things I still don't understand with KG's diagram. I didn't notice the tesuji that means  is big. The tesuji is black can play two spaces above , that gains around 1 1/3 in sente or 4 in gote with a crude atari. The fact there is this big move breaks my argument already.I didn't notice black could clamp in the upper left gaining around 1 2/3 in sente locally, but a bit less as it concerned the ko to the right of it. White connecting in the centre is a 2 point reverse sente, so isn't immediately clear why is better (or perhaps it is just a matter of style and both are fine).The descent in the lower right is also big, and as far as I can tell black hane is 3 to 4 points in gote. I guess there isn't much point for black to take a reverse sente when white can also take a reverse sente, so my logic for my diagram was wrong. E5 still seems hard to evaluate. Because even after the   exchange in your diagram, the next move is still pretty big, 3 points gote at E6.
|
|
| Author: | thirdfogie [ Sun Apr 18, 2021 2:18 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
This post is to tidy up some loose ends and act as a reference if I ever need to revisit the topics. dhu63 Quote: I didn't notice the tesuji that means is big. The tesuji is black can playtwo spaces above My opponent found during the game. He has started looking for moves likethis on the first line. John Fairbairn Quote: I used Kombilo to check the position of two different-coloured ikken-tobis side by side with one side wedging between the opposing ikken. Kombilo came up with lots of examples in the 110,000-game GoGo database - and the inside and outside ataris as played by pros came up exactly even! Thanks for checking with Kombilo. Your results were salutary. One difference is that my opponent plays the wedge when it is a bad move. The professional-standard moves you examined would presumably not be giving the opponent a chance to make a big gain. Quote: The reason this particular post of yours interests me is that you say someone has been telling you that passive play is to be avoided. But who? I've racked my brains but I can't recall the word passive being used in oriental texts. Very occasionally there are words that tend in that direction (e.g. "negative"), but these are very occasional, and my gut feeling is that such comments are applied to beginners far, far more than to stronger players. You perhaps inferred too much. My word "simplistically" should have been "idiotically". The kind of advice I had in mind is "Do not follow the opponent around the board" and "If the opponent plays a so-called double-sente move, look for a similar move you can play elsewhere". No doubt both rules are too simple as stated, but I have certainly found them useful. I cannot read oriental texts, so cannot quote original professional-level material if that is what you were asking for. Yoon YoungSun does have a recent post (https://www.youtube.com/watch?v=8LrJF6huDso) with the strap-line "Effective and Aggressive Haengma", where she discusses responses to an opponent's peep. She is really explaining efficient responses, but the word "aggressive" is there, alongside a cartoon crocodile. Be that as it may, John's reply prompted a wider survey on my side. A search of 64 older games against the same opponent were found to contain 26 wedge moves by him and 7 by me. Details of the search are hidden below. In the 64 games examined, there were 26 wedge moves by him and 7 by me. All but 2 wedges were bad according to KataGo. It would be too much work to make a diagram for all of them, and one should not spam the forum with examples of bad play. Excruciating details are hidden below. Of the 23 non-obvious cases where my opponent wedged, 7 of my replies were wrong and 16 were right. One reason I did better against this collection of wedges than in the original 4, is that 7 wedges in the second set happened in shapes that are familiar from joseki or middle-game joseki, and all my replies were correct in those cases. (I know very few joseki but my opponent makes a virtue of knowing none.) Of the same 23 cases, the "outside" move was correct in 13 cases, the "inside" move was correct in 9 cases. In the other 2 cases, it was not possible to say which side was "outside". Taken together with the 4 cases in the original post, there may yet be some value in my original idea that "playing the outside atari" is often better against this particular opponent. My idea of "outside" is a vague blend of "towards the centre of the board" and "away from the opponent's strength". In the 7 cases where I wedged, 2 were correct. In 2 of the other 5, the best move for white was actually a clamp. That discovery alone justifies the exercise. I don't regret spending time on the research, but it has now been overtaken by events. My opponent has access to these findings and has stopped playing wedges. He now favours lots of invasions on the first line, which are easier to read out. Meanwhile, I am blundering away won games by not counting liberties when tired near the end of the game. It is unclear how to cure that ill. |
|
| Author: | dust [ Sun Apr 18, 2021 5:18 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
thirdfogie wrote: I don't regret spending time on the research, but it has now been overtaken by events. My opponent has access to these findings and has stopped playing wedges. He now favours lots of invasions on the first line, which are easier to read out. Meanwhile, I am blundering away won games by not counting liberties when tired near the end of the game. It is unclear how to cure that ill. Maybe you need to retrain your neural network to play so as to win games early and decisively. |
|
| Author: | jlt [ Sun Apr 18, 2021 5:39 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Even pros are not immune to that disease: https://lifein19x19.com/viewtopic.php?p=233171#p233171 In principle, one way to avoid that kind of accident is, at every stage of the endgame, read what happens when the opponent pushes in every hole, makes all possible ataris and fills all liberties. But that's theory... |
|
| Author: | Bill Spight [ Sun Apr 18, 2021 6:02 pm ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
thirdfogie wrote: In the 7 cases where I wedged, 2 were correct. In 2 of the other 5, the best move for white was actually a clamp. That discovery alone justifies the exercise. Very good.  thirdfogie wrote: I don't regret spending time on the research, but it has now been overtaken by events. My opponent has access to these findings and has stopped playing wedges. He now favours lots of invasions on the first line, which are easier to read out. Meanwhile, I am blundering away won games by not counting liberties when tired near the end of the game. It is unclear how to cure that ill. It doesn't hurt to start with the straightforward approach. Always count liberties. Even if that means doing less reading in the early and middle game. As a result you may make worse plays from time to time. But you won't throw away games from miscounting liberties late in the game. My guess is that you will win more games. In addition, practice makes perfect. Your skill at counting liberties will improve, and will probably take less time as you get better at it.
|
|
| Author: | dhu163 [ Tue Jul 13, 2021 4:45 pm ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Just had some thoughts on Uberdude's point about passive play. My thoughts are about defining such human concepts in terms of shape. A definition of "passive" vs aggressive (or vs active) is probably about taking less risk with your own groups by playing close to them for safety, prioritising reinforcement as opposed to impact on the opponent or unoccupied areas). It is fairly independent of the component of quality/size of a move. But when it is a criticism, it generally means the opponent profits from it, either by taking a bigger move globally (since you didn't need to defend yet) or taking advantage locally. This case can be said to be that the descent allows black to seal off the upper right in sente, so the exchange seems timid. Here I think thirdfogie describes the capture as passive since it removes the chance of the cutting stone escaping. Whereas descending plays more close to Black's area, which seems more confrontational at first sight. However, with some reading (since black playing on White's liberty shortage of L16 seems much less likely that using the N18 sente immediately), there is no chance of the cutting stone escaping if white plays descend. In fact, compared to capturing, descending adds a second buffer zone between the cutting stone and Black's upper right (other than the capture move) making it more difficult for black to encroach. Unfortunately this protection comes at a heavy price, namely the cost of another move given that white is forced to capture the stone anyway pretty much immediately. Uberdude prefers to say the direct capture is not as passive. This is because the main line risks a ko, simply because Black's atari on the ko shape is not sente. I'm sure this distinction between types of passiveness recurs throughout Go. It is related to concepts of gote no sente or sente no gote, or whether to defend a shape with one move or two. . Can this sort of analysis be done with other common words, such as light/heavy? |
|
| Author: | iopq [ Sun Aug 08, 2021 5:04 pm ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
dfan wrote: thirdfogie wrote: There is not enough material to draw any general conclusions. However, accurate reading always helps, and one should probably capture any cutting stone on the third line immediately if possible. In these sorts of atari situations, I believe that in general stronger (let's say 1d) players are much more likely to just capture a stone, and weaker (let's say 5k) players are more likely to extend instead of taking the stone off the board, which looks more efficient but often leads to the opponent getting forcing moves later or having some aji still around. I think that it is a good instinct for players of our level to first consider making the capture, unless there is clearly zero aji. The AI often plays on the 6th or 7th line instead of capturing a stone in a ladder because capturing is over concentrated |
|
| Author: | dfan [ Mon Aug 09, 2021 5:31 am ] |
| Post subject: | Re: Retraining the Neural Network in My Head - Part 2 |
Yes, ladders are another story. All of my comments were referring to the decision between capturing a stone on the third line (by playing on the second line) or descending to the second line. |
|
| Page 1 of 1 | All times are UTC - 8 hours [ DST ] |
| Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |
|
sequence is questionable as white can use black's shortage of liberties around E5 instead.)