Life In 19x19 http://www.lifein19x19.com/ |
|
This 'n' that http://www.lifein19x19.com/viewtopic.php?f=12&t=12327 |
Page 19 of 53 |
Author: | Bill Spight [ Fri Jul 21, 2017 12:05 pm ] |
Post subject: | Re: This 'n' that |
Well, it has been a while, and this may be a slight digression, but it is about AlphaGo. I'll get back to my own theme later. Uberdude wrote: AlphaGo seems to have a tendency to flit around the board, making exchanges at what may appear random times. Is there some meaning behind these moves: in that they are probes and depending on how one is answered play will resume at another place in a different way, or are they just Monte-Carlo style random exchanges which could have been made another time? This quote is from Uberdude's study journal ( viewtopic.php?p=221574#p221574 ) commenting on AlphaGo self-play game 38. I don't think that we can attribute that tendency to Monte Carlo randomness. Pure Monte Carlo sucks. Monte Carlo Tree Search (MCTS) is very good, as we know. But I think — and I would be happy to be corrected — that the combination of Neural Networks with MCTS is a hybrid. The neural nets are not just modifications of MCTS to make it better, they are a different approach, and each approach gains from the other. This "flitting about", I suspect, has two main reasons. First, AlphaGo does not make plans. Yes, it builds a search tree, but that is not the same thing. Znosko-Borovsky advised humans to make plans (in chess) and not to try to try to find the best play. AlphaGo tries to find the best play. It may randomize among plays that have nearly the same evaluations, but its evaluation is based in part on randomness, so maybe not. (Uberdude, maybe this is what you had in mind with Monte Carlo randomness. ) But even if its evaluation is not random at all, I think that we would get the same effect through errors. Any evaluation function in the opening will be uncertain. Second, I think that the tenukis reflect the nature of the opening. There are relatively few sente in the opening. Let me give one of my favorite illustrations. At the beginning of modern thinking about the opening, a few centuries ago, Black would play on a 3-4 point, White would make the 5-3 approach, and Black would make a pincer, usually one or two spaces. Then White would play in a different corner. This was actually a pretty sophisticated idea by White. White could be satisfied with having prevented an enclosure with sente. After a while, Black stopped making the pincer, but played in a different corner himself. The approach was not sente. Eventually White stopped playing the approach on move 2, although it has not entirely died out. Consider the situation after the Black pincer at move 3. The pincered stone has no base and is subject to further attack. Doesn't protecting it have some urgency? In the current vernacular, isn't that corner hotter than an open corner? The answer is no, and the ancients knew it. So does AlphaGo. As I have explained before, in general, as stones are added to a region of the board, the local temperature drops. OC, there are sente, and the local temperature might not drop if the stones were played randomly, but they are not. Stones of the same color are played to help each other, or to work together. This coordination tends to strengthen the stones and thus to lower the temperature. There are hotter plays elsewhere. Now, this is something I have known for quite some time, hence my "proverb", Tenuki is always an option. But AlphaGo has surprised me. For instance, The stones are strong, and so I have believed that the local temperature has dropped, and often tenuki. However, AlphaGo usually does not tenuki now, but connects the stones. AlphaGo usually makes the solid connection and tenukis after . But the stones do not have a base, nor do they have any eye shape. Don't they need to extend to "a" or something? Apparently not. (Although sometimes AlphaGo does make an extension.) This sequence, playing and then tenuki, shows that AlphaGo's tenukis are not the result of randomness. If we find them hard to understand, I submit that the fault, dear Brutus, is in ourselves. |
Author: | Bill Spight [ Thu Jul 27, 2017 8:26 am ] |
Post subject: | Re: This 'n' that |
Imitating AlphaGo There are those who believe that imitating AlphaGo is not worth much to human players. I disagree. Humans are very good at imitation. And imitation is an important way of human learning. As the saying goes, most of what we learn is caught, not taught. Let me say a bit about imitation. Imitation is learning through observation. For a long time in learning theory imitation took a back seat to conditioning. In operant conditioning behaviors are reinforced or extinguished through reward and punishment (in both literal and metaphorical senses). Much of training is still based upon operant conditioning, using it to shape behavior. At one time behaviorists even thought that humans learned language through operant conditioning, that babies made random noises (babbling) that parents shaped by reward. We now know that babies' babbling is not random, but is imitative. (And not everybody agreed with the behaviorists, anyway. ) A great deal of language learning is, in fact, imitative. Children learn the rules of grammar, which can be quite complex, by imitation. True, schools teach grammar, but they teach standard grammar, in an attempt — not always successful — to correct the grammar that children have already learned. In addition, children pick up thousands of words per year, without much in the way of instruction or reward and punishment. I bring up language learning to emphasize that imitation can be cognitively sophisticated, not just monkey see monkey do. Imitation got a boost in regard in the latter half of the twentieth century by the demonstration that some learning by other species was not necessarily by imitation. For instance, back in the day when milk was delivered to back porches, some birds in England learned to peck off or peck through the milk bottle caps to get at the milk or cream on top. This was widely regarded as an example of imitation. But one possibility is that the birds did not imitate other birds, but simply learned that the tops of milk bottles might be a place to find food. This reframing of learning in other species came about because of the recognition that imitation is cognitively advanced. Anyway, some imitation is of the monkey see monkey do variety, and there is a go proverb that warns against such imitation: Learn joseki and lose two stones in strength. I certainly followed that proverb, along with the advice of Znosko-Borovsky not to make opening moves without reflection. When, as a dan player, I did study joseki, it was not to memorize them, but to understand them (as well as I could, OC ). But I have now come to appreciate both memorization and even monkey see monkey do imitation, especially in the early years of learning go. Most of what we learn is caught, not taught. In my case, as an SDK I studied Go Seigen's games and imitated him, not in the sense of Go Seigen made this play so I will too, but in the sense of trying to play like Go Seigen. Some people warn against that, because you will never understand go like Go Seigen. Better to play moves that you understand, by which they mean consciously. It is true that you or I will never understand go like Go Seigen, but people grow in understanding. Children play pretend, and that is a good learning strategy based upon imitation. It works. As aspiring pros imitate AlphaGo and other superhuman programs, as they come along, I think that we will see a rapid advance in the development of go. I say so not because AlphaGo, et al., are or will be good at the calculation of variations, because humans are not very good at that, but because those programs are and will be good at strategy and move selection, which humans can imitate. And humans are very good at imitation. |
Author: | John Fairbairn [ Thu Jul 27, 2017 9:44 am ] |
Post subject: | Re: This 'n' that |
Bill Good points, and I agree with you completely, not least because I have literally just seen my latest (6-month-old) grandchild learn by imitation. A couple of days ago it was learning to wave goodbye, and today he was imitating me, cocking his head first one way then the other. On top of that, someone mentioned "Zen and the Art of Archery" to me a few days ago and that prompted me to re-read it. It confirms what you said. For those who haven't read it, it is the tale of a westerner (a university philosophy lecturer of the rational German school) who was keen to learn the "theory" behind those Oriental arts that had a (to westerners) mysterious element. He was lucky to get a posting to Tokyo and chose archery to study as a Japanese colleague of his also did it. But the Japanese master initially refused to take him on because he had once before accepted a foreigner and now regretted it bitterly. But he was eventually persuaded by the Japanese pupil of his, who effectively stood guarantor. The philosopher was first given the task of practising extending the bow, as demonstrated by the master, and tried for weeks without any success. He finally lost patience. Only at that point did the master tell him how to do it. But it wasn't a trick, as the German had expected. It was simply that the student had been imitating the wrong things. He had not been imitating the master's breathing, just the arms. Initially the German student was angry. Why couldn't the master tell him that at the beginning and spare him all that frustration? But the Japanese student pointed out that the master understood the pupil as well as the art of archery, and knew that he had to go through a "babbling" phase before he could imitate properly, and then understand. If he had taught the student to imitate the breathing at the outset, the student would only have learned to imitate that (be taught) and never to understand (catch) the deeper aspects. maybe we can view the 60-0 trouncing AlphaGo gave the pros as reducing them to babbling? They have already started imitating. But are they imitating the right things? Regrettably the AI equivalent of the master archer has yet to be "imitated" by AI, but that means there is a still a lot of fun to be had in discovering go's mysteries. |
Author: | Bill Spight [ Thu Jul 27, 2017 1:40 pm ] |
Post subject: | Re: This 'n' that |
Imitating AlphaGo: Self play game 7 I want to explore the question of imitating AlphaGo through looking at one or more of its self play games. Some thoughts about game 7. is an AlphaGo special, a side attachment on the fourth line against a high enclosure. An eminently imitable play. Once you've seen it, you can grok it. Let's do a little grokking. is a standard approach to this enclosure. However, it invites , a good play combining extension and pincer. I don't like it. My first thought on the bottom side would be this pincer. However, is a good extension and counter-pincer. The attachment, OTOH, invites a hane or a nobi. It is not a territorial play, nor does it create thickness. But so what? It creates aji. Another attractive play for White is an extension on the right side, or "a" or "b", especially as Black threatens a pincer cum extension there. may look a little passive, but it avoids strengthening the stone. Next, plays the standard attachment to this enclosure on the right side. - is the standard sequence in reply. If is at 15, White can play the hane at "a", now or later, which would work nicely with . Note that if White started with at 12, and played the side attachment at 10 after Black had strengthened the corner with 13 and 15, Black would not now play the descent at 11. is the desired extension on the right side. If White had played it before , Black would have had a nice extension to "b". But the aji of and make "b" unattractive now. Edit: BTW, I think that extending to "c", my original thought for the side extension, would not be so good, precisely because it forms a base with and , instead of treating them as aji. I think that would be heavy play. Now, my analysis is not imitation. I am not learning by observation alone, I am hypothesizing about AlphaGo's play. But what I am saying is something that a player might just pick up from AlphaGo's play, without giving it much thought, especially a kid. From this and other examples I think that one way to play like AlphaGo, to imitate it, is to play probes and create aji. This is different from making bases or territory or creating thickness or large frameworks. And, OC, just because AlphaGo makes certain plays or plays a certain style does not mean that other plays or styles are not just as good. Go is hardly solved, especially in the opening. Anyway, I am not the first to notice that AlphaGo plays probes and creates aji. More on this game later. |
Author: | Bill Spight [ Thu Jul 27, 2017 1:52 pm ] |
Post subject: | Re: This 'n' that |
Thanks, John. Zen and the Art of Archery. An excellent book! Thanks for reminding me of it. |
Author: | Cassandra [ Thu Jul 27, 2017 2:36 pm ] |
Post subject: | Re: This 'n' that |
John Fairbairn wrote: It was simply that the student had been imitating the wrong things. Had the student really been imitating the "wrong" things ? If I remember the lecture of the book correct, the real problem was imitating the "right" things (i.e. "technical issues"), but conciously. And this was by far not enough, but seen by the master as "cheating". I think that the key was totally forgetting about the "techniques" and acting unconciously. |
Author: | moha [ Fri Jul 28, 2017 6:58 am ] |
Post subject: | Re: This 'n' that |
Bill Spight wrote: As aspiring pros imitate AlphaGo and other superhuman programs, as they come along, I think that we will see a rapid advance in the development of go. I say so not because AlphaGo, et al., are or will be good at the calculation of variations, because humans are not very good at that, but because those programs are and will be good at strategy and move selection, which humans can imitate. And humans are very good at imitation. I agree about the advance of development, but disagree on the rest. First, I'm not at all convinced that "strategy" is a particular strength of AI. I get a different feeling from some of these selfplay games. It seemed that sometimes it chooses a line which disagrees with human strategy, but it seems to think it has a good enough line to keep its game together. Sometimes this works, but sometimes this is a horizont effect: the bad consequences come slowly and finally the side that went against strategy falls behind. Even if AlphaGo would still play strategically, consider tomorrow's AI. Sooner or later we will have near-perfect programs that read very, very deep. Strategy is meaningless at that levels, only calculation of variations remains. So humans will find themselves in an environment discussed earlier (Kasparov example): they can (use an AI to) look up the best move at any position, but no explanation (beyond a minimax line). I think strategy and any such explanation will be left for humans to formulate. Strategy is about probabilities, the human way of dealing with the uncertainity (unread lines). But we WILL also see what is the percentage for cases where all attempts of strategic explanation fail, and a move is good ONLY because a good minimax line happen to exist. |
Author: | Bill Spight [ Fri Jul 28, 2017 10:00 am ] |
Post subject: | Re: This 'n' that |
moha wrote: Bill Spight wrote: As aspiring pros imitate AlphaGo and other superhuman programs, as they come along, I think that we will see a rapid advance in the development of go. I say so not because AlphaGo, et al., are or will be good at the calculation of variations, because humans are not very good at that, but because those programs are and will be good at strategy and move selection, which humans can imitate. And humans are very good at imitation. I agree about the advance of development, but disagree on the rest. First, I'm not at all convinced that "strategy" is a particular strength of AI. But strategy is what allowed AlphaGo to leap ahead of pure MCTS programs, and humans, as well. The neural nets are strategic in nature. Quote: I get a different feeling from some of these selfplay games. It seemed that sometimes it chooses a line which disagrees with human strategy, but it seems to think it has a good enough line to keep its game together. Sometimes this works, but sometimes this is a horizont effect: the bad consequences come slowly and finally the side that went against strategy falls behind. The horizon effect is not the only source of errors. Building a search tree is not the distinguishing characteristic of AlphaGo, or even MCTS systems. Although the latter use an improved algorithm for building the tree. Quote: Even if AlphaGo would still play strategically, consider tomorrow's AI. Sooner or later we will have near-perfect programs that read very, very deep. Once upon a time, chess programs tried to emulate human play. But programmers did not know how to do that very well. What they did know how to do was search, and eventually the combination of hardware and search algorithms produced world class programs. At the same time (the 1990s) the same search based approach produced amateur kyu level programs in go. The advances in go programs since then have not come about from the same approach. Future advances to computer go may rely upon increased depth of the search tree, but I kind of doubt it. Quote: Strategy is meaningless at that levels, only calculation of variations remains. Even tic-tac-toe, which is completely solved, has strategy. |
Author: | Bill Spight [ Fri Jul 28, 2017 11:17 am ] |
Post subject: | Re: This 'n' that |
Imitating AlphaGo: Self play game 7, continued is another AlphaGo special, the early, unsupported 3-3 invasion. Why did White block on the top instead of the left, I don't know. AlphaGo imitators may go either way. The keima, , is AlphaGo's usual choice, unlike humans up till now, who usually hane. occurs in human play, but "a" was by far the usual choice. Is superior to "a"? No, we can't say that. For one thing, AlphaGo has played a few more recent games against itself with longer time limits, spending 120 sec./move instead of 45 sec./move, as in this game. In those games AlphaGo has chosen "a". (OC, we cannot conclude that "a" is superior, either. ) completes one of the AlphaGo joseki in the top left corner. Note that makes the ladder work. Is that why White chose to block on the top side? ¿Quién sabe? The next two plays are quite ordinary. stakes out the bottom side, and plays at the frontier of opposing frameworks, enlarging the White framework while reducing the Black framework. Wa and Wb are also possible, but AlphaGo tends to prefer higher and thinner options. AlphaGo often plays the attachment, , with the and stones in place. Here the strategy of strengthening the edge of one's framework is well known. With Black strengthened, White can be glad that is not closer to the Black strength. Also, works well with . is not unusual, but to me it does have that Go Seigen - AlphaGo flavor. also has an AlphaGo flavor, being a high shoulder blow in preference to other possible reductions. Note that AlphaGo seems unconcerned that White builds a huge framework. A lot of top pros share such unconcern. FWIW, I would have wanted to play a reduction earlier, such as at 42 or "a". Maybe that would not have been so good. {shrug} Anyway, if I were trying to imitate AlphaGo I would cultivate that unconcern. enlarges the Black framework while reducing the White framework. Note that after White does not defend at "a" or "b", but jumps to instead, making use of the aji of . There is a good lesson in light play there, I think. White threatens to push through at "c", sooner or later. Black attacks with the ogeima, . There is a saying, Keima for attack. Characteristically, AlphaGo plays one point higher and one point thinner. is, I suppose, the choice of a plurality of pros. Attachment for sabaki. Amateurs might respond at "d", but again, I think that is the pro choice. The subsequent fight is quite interesting, but I cannot identify anything peculiar to AlphaGo in it. |
Author: | moha [ Fri Jul 28, 2017 11:34 am ] |
Post subject: | Re: This 'n' that |
Bill Spight wrote: moha wrote: First, I'm not at all convinced that "strategy" is a particular strength of AI. But strategy is what allowed AlphaGo to leap ahead of pure MCTS programs, and humans, as well. I still feel that the NN is more useful for pruning the search, than for drawing conclusions by itself. It effectively allows the program to do USEFUL search for the first time. As for humans, I think they have more than enough strategy to compete with Alphago - they lost because of their weak reading ("feeling is not enough, you need to read everything" = global minimaxing). Quote: Quote: Even if AlphaGo would still play strategically, consider tomorrow's AI. Sooner or later we will have near-perfect programs that read very, very deep. ... At the same time (the 1990s) the same search based approach produced amateur kyu level programs in go. The advances in go programs since then have not come about from the same approach. Future advances to computer go may rely upon increased depth of the search tree, but I kind of doubt it. Sooner or later - after QC at the latest. But strategy is inferior to reading in the end. Quote: Quote: Strategy is meaningless at that levels, only calculation of variations remains. Even tic-tac-toe, which is completely solved, has strategy. It has strategy, but that is meaningful only when you do not read to the end. I think "strategy" itself is the collection of methods and theories for dealing with unread lines - or interpreting the completed solution, to draw conclusions for future similar positions with unread lines. |
Author: | Bill Spight [ Wed Aug 23, 2017 8:20 am ] |
Post subject: | Re: This 'n' that |
Three Points Without Capturing Yeah, you were expecting something else. This one. The one with the possible ko. It is unfortunate that Shuwa did not explain the reasoning behind his ruling. As for the ko, it is plain that, as with Bent Four in the Corner, he assumed that there were no ko threats. However, there is still the question of why Shuwa said that Black does not have to capture the four stones. (She does in the current Nihon Kiin rules.) The reason, I think, is the same as why Black does not have to capture the single stone in the first diagram. White has no play to gain anything. In the first diagram White has no play at all, suicide not being allowed, and in the second diagram the result is three points for Black, and more if Black can win the ko. (OC, if suicide were allowed, the result of the suicide would be four points for Black.) To put it another way, the capture of the White stone or stones would take place at temperature -1, where each play costs one point. OC, Shuwa did not have the modern concept of temperature, but by his ruling alone, he recognized that the capture would lose a point. |
Author: | John Fairbairn [ Wed Aug 23, 2017 8:56 am ] |
Post subject: | Re: This 'n' that |
Just to be pedantic, Bill, we have no record that Shuwa, or anyone else, made any such ruling. There is a legend that a position was referred to him for arbitration but this seems likely to be a conflation with the fact that a gomoku torazu occurred in one of his games. But it was not actually played out as such and so there was no need for a ruling. More than likely there was a discussion about it after the game and just as likely Shuwa expressed an opinion. But we don't even know what his opinion was (unless you infer it from the name, and assume he created it - but as I say his position was gomoku and not sanmoku) let alone the reasoning. There has been no need in pro practice for a ruling since as far as I know. |
Author: | Bill Spight [ Wed Aug 23, 2017 9:49 am ] |
Post subject: | Re: This 'n' that |
John Fairbairn wrote: Just to be pedantic, Bill, we have no record that Shuwa, or anyone else, made any such ruling. There is a legend that a position was referred to him for arbitration but this seems likely to be a conflation with the fact that a gomoku torazu occurred in one of his games. But it was not actually played out as such and so there was no need for a ruling. Thanks for the correction, John. If the torazu gomoku was not questioned, then torazu sanmoku was probably already accepted. BTW, which game of Shuwa's was that? Thanks. |
Author: | John Fairbairn [ Wed Aug 23, 2017 11:18 am ] |
Post subject: | Re: This 'n' that |
Quote: BTW, which game of Shuwa's was that? Versus Sanchi in 1839. |
Author: | Bill Spight [ Wed Aug 23, 2017 1:31 pm ] |
Post subject: | Re: This 'n' that |
John Fairbairn wrote: Quote: BTW, which game of Shuwa's was that? Versus Sanchi in 1839. Much grass. |
Author: | Bill Spight [ Fri Aug 25, 2017 7:59 am ] |
Post subject: | Re: This 'n' that |
Here is a problem based on the weiqitv commentary of the recent Wang Haoyang vs DeepZen game in MLily Cup. See https://www.lifein19x19.com/forum/viewt ... 86#p220886 . and each gain 2.5 pts. is correct, as it gives White a good follow-up, , which gains 2 pts. The temperature has dropped to 1 pt. Black to play and get the last play. |
Author: | Bill Spight [ Fri Aug 25, 2017 9:32 am ] |
Post subject: | Re: This 'n' that |
Here is the actual game from a similar position. Earlier in the game (B 215), Black made the exchange, - , which was a losing sente that sacrificed 3/4 pt. Peng Quan, the commentator, thought that that was the losing move. B 231 sacrifices an additional 1/6 pt. to get B 233, which gains 2 pts. Despite the fact that White plays first at temperature 1, Black can get the last play. So it turns out that B 233 is the last effective play of the game. Gaining 2 pts. with the last play may well justify B 215. White took his 1 pt. sente, and then both players erred. W 238 is a mistake, as was B 239. In this position, Black to play and win. Edit: Sorry, White can still win. I made a miscalculation. |
Author: | Schachus [ Fri Aug 25, 2017 11:47 am ] |
Post subject: | Re: This 'n' that |
Maybe I'm wrong, but: So my answer is Edit: Of coarse what I wrote is for japanese rules, in chinese, this might change |
Author: | Bill Spight [ Fri Aug 25, 2017 12:03 pm ] |
Post subject: | Re: This 'n' that |
Schachus wrote: Maybe I'm wrong, but: So my answer is Edit: Of coarse what I wrote is for japanese rules, in chinese, this might change Since there are no ko fights to speak of, nor seki, correct play by Japanese rules is correct by Chinese rules. In reply to your edit: |
Author: | Schachus [ Fri Aug 25, 2017 12:13 pm ] |
Post subject: | Re: This 'n' that |
Page 19 of 53 | All times are UTC - 8 hours [ DST ] |
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |