Appreciating AlphaGo
Vložit
- čas přidán 19. 01. 2017
- I strongly believe that amateurs can appreciate and learn from AlphaGo's games, and in this video I try to show how by looking at Lian Xiao's first game against AlphaGo.
Thanks to Jennie Shen 2p for providing variations and professional thought processes for the video. Don't blame her if I got something wrong though. Any mistakes are 100% my own. - Hry
![[Life&death] #1 God of Baduk Lee-Changho 9dan professional player even couldn't solve this](/img/n.gif)
I have only recently discovered your channel after rekindling my enjoyment of watching Go (although i have yet to take the plunge into playing)
After watching a bunch of Nick Sibicky videos your name popped up ... and you just have such a positive and relaxing view on the game... please keep up the videos! :)
+1 to this!
+1
Agree, I really like your videos and reviewing of these games. I am a KGS 1D from 2009 so I can follow most of the things your explaning. I have yet to start playing again but who knows! I still like the game whenever I see it. And AlphaGo is very interesting to follow and learn from.
I do love your AlphaGo reviews, and I'm very happy that you offer just the right amount of background explanation and reading out of moves so that people as amateur as me can enjoy these games. ;w;
AS A SERIOUS BEGINNING GO STUDENT YOUR VIDEOS ARE MORE HELPFUL THAN ANY
OTHERS....THANX
I like it when you impersonate AlphaGo's insight's: "Oops! now you're a stick!" "Now that stone would look stupid" it cracks me up. It also help to understand moves and motives
At 35:56 it's a bamboo joint not a tiger's mouth.Just a slip of the tongue of course, but I mention this because I have the feeling that Alphago uses this form a lot. Or is it just my imagination?
I've been looking for a simple Go commentary that I can understand for a while. Everything I've found isn't really geared toward beginners or helping low skills. You make things simple I really like that.
You are asking whether we want to see more of your videos. This is a silly question Brady. Your videos are awesome!
Thank you for posting these! I played Go about 10 years ago quite frequently and was something like 8 kyu before I stopped, but have followed it from time to time. AlphaGo instantly got my attention and it is really educative and enjoying to watch such high quality commentaries from the latest games. Kudos!
Huge thanks from in-my-best-times 13 kyu. Without your commentary, I'd be lost and fail to see the beauty of it!
Thank you for all the videos they make, they are helpful for a beginner such as myself! They're all interesting and easy to understand.
I also greatly enjoy these AlhpaGo videos, and hope to see more of them in the future!
I don't play Go at all but I watch your videos because of the way you talk. Keep it up!
Thank you Brady, keep doing what you're doing! Really appreciate your breakdown of AlphaGo games for us fellow amateurs to enjoy.
Really enjoyed the clear analysis, thank you. More please!
Great video Brady! Nice, provocative insights.
Awesome! Can't wait for the next AlphaGo video.
Thanks, Brady. Love you, Jennie!
I hope you'll be continuing with this series. It's quite interesting!
Love the AlphaGo videos and yes, would like to see more. AlphaGo v. AlphaGo would be interesting to see commentary on if it's not too complicated.
Awesome you did another Alphago video ^^
Thanks for making another AlphaGo video. Keep it up :)
I love your videos so much. Please record AlphaGo's games in late may for us!
Awesome. We're living in crazy times of history
great review, thank you!
Hi Brady,
I found and subscribed to your channel today after watching your analyses of two of AlphaGo's games. I really liked your videos profiling the strengths of AlphaGo and what we can learn from it. In fact, we will have to go through this process if we are ever going to win a game against AlphaGo.
--
In fact, as you pointed out, AlphaGo hasn't really shown its fighting skills. So, it has effectively declared that it wasn't challenged at all in all those games where it was able to simplify the game as you put so well in an example.
--
Watching your videos, I am quite charmed with AlphaGo's skills and am looking forward to learning more from it. Also, I am curious to know if there were any pros that have indeed managed to challenge AlphaGo into fighting. If so, could you show such a game? It might be closer to a won game for its opponent.
--
Alternatively, I am quite happy to listen to your analysis of many more of AlphaGo's games in so far as they are interesting to you in any way.
--
I am looking forward to your video with Jennie Shen, when you have all the equipment to make it work.
--
I had one question, where you got the game info about AlphaGo's games.
--
Thank you very much.
Thank you Brady!
I've written a minimax implementation for a two player zero sum game so perhaps my insights would be of some value here. It was a simple algorithm that read 10 to 12 moves ahead and picked the biggest move. Its play displayed a similar flexibility to what we see in alpha go. As the board changed it seemed to come out on top in trades. However as I examined its play it became obvious it sometimes made poor moves. A few moves deeper it, if it was playing against a strong opponent, it would realize it made a mistake and would try to recover. Due to the ridiculously large size of the game tree and its good reading ability it would usually be able to accomplish this with only one or two point loss all the while making it look like that was its plan all along. And that was when it played against very strong players. Against slightly weaker players all of its moves looked good because mistakes didn't get punished. I would argue that that's what we're seeing with alpha go. I doubt alpha go has as much certainty about a move like E17 as you imply. In its other games as well as this one I see alpha go get itself into a pickle with overplays where it creates too many weak groups and then skillfully mitigates most of the damage. The pros aren't able to effwctively punish its overplays due to their weaker reading ability.
I enjoyed it a lot! I aM waiting for More AlphaGo gaMe froM you.
please make more AlphaGo games! Really nice thoughts, helped me personally a lot, i would love to see you go over some alphago vs alphago games - the "god" of go still loses, if he plays another god of go
Of all the many AlphaGo reviews, I like yours most
great channel!
Hi brady, I just found you channel, and I think it is great. I really like your style, the humility is very fresh and I think very courageous of you. Sharing blunders is not easy.
I wanted to ask you: would you mind creating a chronological playlist? I would like to watch your videos from the beginning.
It could be because you're just a good teacher and analyst, but as an amateur chess player who barely knows any Go theory in the days when she tried it, I understood everything you were talking about! I noticed there is less of a intuitiveness when it comes to a similar style analysis in chess. Chess can't, compared to Go, be as understood as a game through pure big-picture vantage points. Nitty-gritty details are needed; you have to evaluate positions closer to its reality than your opponent to win a game at all its stages. There doesn't seem to be such a need in Go though, and that's what makes Go beautiful.
thanks!
Some questions and reflections from an amateur: The first one is that black's third move (the 5th move in the game) is the wrong extension for a reason -- and AlphaGo punished it immediately and forever in the game; that stone was always in the wrong place. When learning about the very basics on extensions this one is shown as a big blunder (maybe the pro wanted to surprise the computer with something new, but still...). The second one is that I want to ask if the sequence after white's move on R-10 is forced; isn't black too accommodating, especially after foreseeing the results in the lower right side? Usually pros try to respond more aggressively; couldn't he have run on the fourth line or play at R-8 - create a bigger fight -- instead of enclosing so early? Another question is regarding the lower left sequence: isn't it probably better locally and strategically to jump at F-5 instead of attaching the pincer stone? After the sequence in the game is played everything goes wrong for black in the other two corners; I guess Lian was frustrated and just preparing the scene for a resignation but it seems to me that this was not his best effort (too concerned about AlphaGo's recent trend?)
26:11 your dialogues are very accessible. Thank you.
Great video, thanks. Did you have that game memorised?
I'm curious if they've allowed AlphaGo a do-over on the Lee Sedol game that AlphaGo lost, and what it would have changed about it.
Thanks for all the reviews, but what is the program you use for reviewing them? For showing the board and the variations etc.
Very interesting. In the what if category. What if, they play on a much bigger go ban., let's say 101 X 101. From what i know, that jump on size would be unmanageable for Alphago, many years. On a other hand, expert players can be ready to play quite fast. And so on with increasing size.
i would like to see more brady
Brady,
On fighting: It's a serious military maxim that to keep the peace you have to be ready to fight. We tend to ignore this because we usually hear it from the mouths of rightwing blowhards trying to get more taxpayers' money for their friendly local defence contractor. It is, however, once you get it surrounded with all sorts of conditions and context, a large partial truth.
As you've found, AlphaGo gets into fights playing itself. Against humans, it seems to me, it is up against people with deep judgement who know the threats and take them into account.
In real life strategy, there's a parallel. Herman Kahn, the great RAND Corporation strategist, used to talk a lot about World Wars Three and Four, wars which had not, in fact, occurred in real life because both sides understood the game. It was all played out with feints, bobs, weaves, press releases and academic conferences -- and the actual building of, e.g., the Minuteman III, the British Trident force, the North Atlantic electric fence, and lots more.
There was, in fact, some blood and powder in there, too, Vietnam. This, it seems to me, was a bunch of mistakes on the part of the Americans, and it is to America's credit that it eventually figured this out. Most of the "fighting" for the past seventy years, however, has been the sort that you see, or don't see, on AlphaGo's board, the dance of symbolic threats backed up by the knowledge on both sides of what the fights would have been likely to mean.
Cheers (and congratulations on your fine videos),
-dlj.
Hey Good vids!
Where can I download the Kifu for all the games?
17nicksmail Google "AlphaGo 60 games sgf"
Or you could just give the link, like someone on KGS did for me...www.reddit.com/r/baduk/comments/5m0ao4/all_games_30_of_master_p_alphago_from_foxwq/
Oh didn't realize URLs work on CZcams replies now! That link is incomplete actually, there are 60 names, not 30.
Scroll down and read the forum posts...all 60 are there in zipped format.
which go program is this being played out on? Silly question but I'm super uninformed!
Brady is using the cgoban software, the KGS client: www.gokgs.com/
KGS Java client - CGoban
It is called CGoban. Pretty nice since it's lightweight and utilizes Java multiplatform merits
More videos please. Team up with a pro if you can or crib from other sources for comments.
where do you watch alphaGo games?
It is on fox server if you have account. www.reddit.com/r/baduk/comments/4vu0fn/a_guide_to_playing_on_the_fox_go_servers/
You download all 60 kifu at this link www.reddit.com/r/baduk/comments/5m0ao4/all_games_30_of_master_p_alphago_from_foxwq/
make some new videos alphago commentary videos please!
What scares me is they named it Alpha go. Alpha as in the software is still in the alpha stage and is already this good? or is it just a branding name? lol.
I think the choice is one of Alpha as in top dog.
I'm sorry for giving this a bad rating, but when you're showing a play of x vs y, you are NOT supposed to fill over 50% of the video with only you trying out different sequences to the point where nobody knows what was actually played anymore.
I started writing this comment by stating that I strongly disagree with you in that we are more able to understand the AlphaGo game records than pro fights. But after finished writing below, I now take the position that I explain the reasons behind your stance. Also, my comment hopefully functions as the caution to appreciating AlphaGo games.
1. Why do the game records seem to involve much shorter local fights than the pro game records?
I would say that joseki like sequences in AlphaGo AGAINST MUCH WEAKER PROS have become generally much shorter than the previous pro joseki studies like avalanche, since what we have thought "forcing" or "too bad" turned out not to be the case. The subtrees that have been cut by intuition in the past, which resulted in such a long "only play path (ipponmichi in Japanese)", cannot be cut that easily any more. It is only that the current so-called josekis have turned out far from the case, thanks to much better judgment and reading combination by AlphaGo. If you look at the 60 game records, they are like the "study games" among pros testing new endeavors in joseki and other local fights. In a lot of 60 games, we see AlphaGo dominating by far in various local fights. These local fights are simply discarded in the study games and tend not to appear in official game records. The 60 games give us fake simple impression, but they are in fact just like hidden discarded subtrees that appear in joseki textbooks, that are also "simply too bad for one player". Hypothetically, if the pros are allowed to play many many more against AlphaGo, the local fights will become much more sophisticated and will converge into something more like new better josekis (this time, however much more whole board than the previous ones).
If you recall AlphaGo vs AlphaGo game records in summer 2016, I feel the situation changes dramatically. The joseki like sequences may possibly become at least as long and complicated as the previous pro studies. I at least felt it quite difficult to understand these games. I imagine all the more complication in the current version.
2. Evaluation vs.(?) Long Reading
While I agree with you that understanding AlphaGo's moves depends much more on the evaluation than the technical long reading required for understanding pro game records, I should also say, along with several pros, that the attributes you listed in this video are all way too highly integrated in the case of AlphaGo, at least much much much more than what you appear to show in this video. For instance, AlphaGo's positional judgment and clarity can only be validated by her outstanding life and death skills. The overall multi-attribute decision making of moves only seems to be manageable at a high level by the combination of Deep Learning and Monte Carlo. It is not via an additive independent multi-attribute utility function. Without the godlike integration of the attributes, the misjudgment in one or two of the attributes in amateur plays trying to mimic AlphaGo's hyper efficient play style may only end up in disaster. There is in fact a Japanese highest level amateur player tweeting when he tries to mimic tenukis that AlphaGo makes quite often at such delicate timing, he only ends up dying. It is definitely a much longer way for us the more mediocre amateurs.
In the game you presented this time for instance, you attributed the right top sequence to flexibility. However, k17, along with the hidden diagrams involved, is just like the brilliant hazama move in the game three of Lee Sedol challenge match. There is way too much brilliance and reading involved rather than the simple flexibility of possibly sacrificing the top stones involved, such that it looks to me just like another hyper complicated joseki studies made by pros. If we look back old josekis in Japanese texts from the current Chinese and Korean joseki studies, we also have the impression that the old ones are way too narrow minded, inflexible and so forth. In the beginning of this comment, I referred to avalanche as an example, but avalanche has not been played for centuries because the hane on the two stones look too bad a shape. So, rather than simply flexibility, I would say that the top three stones are in fact much less weak than we previously "felt".
The same can be said regarding the early 33 plays by AlphaGo. The outside wall turned out not to be that thick as we "felt", only thanks to the brilliance of no hane-tsugi (at least according to Ohashi Hirofumi's analysis). And even worse, AlphaGo plays 33 at some times while not at other times. It is really extremely hard to tell when is the right timing to play. This is how sophisticated evaluation by AlphaGo is.
I should also comment on your attribute of "preconceived notions".
I think if you use this concept without adequate understanding, and clearly I think you lack one, then you may possibly cause serious misunderstanding.
First of all, you say that AlphaGo does not care about preconceived notions, but I think policy network and value network are both the massive accumulation of "preconceived" notions. Especially the former. Why can AlphaGo play in much less than a second and still probably beat almost all humans 100%? Because it has constructed such high level intuition telling you where it "looks (at first sight)" effective to play via the past massive learning playing against itself. This is exactly like the intuition that humans have regarding the evaluation of the right top fight in the video. Humans, just like AlphaGo, have created their own policy network "feeling" on which they judge the three stones on the top are very difficult to make use of when there is a stone at j17 WITHOUT READING. As you very well know, you definitely need this sort of intuition in Go to cut off "hopefully" unnecessary ineffective branches. Otherwise, you have to study too many variations, which is only possible by computers.
I am sure many amateurs not knowing the effect of j17 have played the small avalanche just like this video, only end up being punished to confirm the "preconceived notion" that you should not head into this sequence when there is an opponent stone at j17. Thus, clearly the "preconceived notions" depend very much on the play level.
I in fact think Lee Sedol and Cho Chikun are very best examples within the human players that have destroyed the preconceived notions in Japanese players. Their brilliant policy net value net have made other players seriously review the concept of "thickness" for instance. AlphaGo just makes us revise our notion of thickness, strength and so forth even more fundamentally.
Again, I assert that you are darn wrong to say that AlphaGo is free from preconceived notions while humans are. We all require fullest use of preconceived notions in the high level play of Go, and we only update the preconceived notions little by little. It is only because AlphaGo has appeared only very very very much once in a while that we have not had the chance to appreciate how AlphaGo has updated its own preconceived notions little by little through the self play. It is only that AlphaGo has created the preconceived notions that far exceed human players' at the current stage.
Actually, the right top is a best example in how humans themselves have worked very hard to fight against preconceived notions. As I said in the previous comment, "don't think to play hane on the head of two stones" is a really extremely effective preconceived notion, which is of course valid still most of the time even after our better understanding of Go than the past. Avalanche joseki was only possible since a very few players have really worked hard to study the variations that feel too bad with this preconceived notion. This sort of adventure pays off only very much once in a while, just like discoveries that make us rewrite laws in science. It is seriously wrong to think that we should free ourselves entirely from the preconceived notions. They are the accumulation of human wisdom. It is only that we now have a better tool, the computers, to accumulate the wisdom more effectively in Go.
decidrophob I'm pretty sure alpha go is making overplays and is able to get away with them due to its stronger reading ability.
I agree with you trucid2 depending on the definition of the term "overplay". The notion resembles the relationship between existence and construction in mathematical theories.
If you are using the terminology in the sense AlphaGo is not playing the GOD moves (the solution of backward induction tree search), then no doubt you are correct. No one still expects AlphaGo anywhere close to perfect plays (though in fact she may possibly be). So, AlphaGo may be playing many overplays taking this definition. We however do not know how much "over" these overplays are. Possibly maximum of a few points or a fraction of point loss.
If on the other hand, you define the term "overplay" only applies when someone actually constructs the concrete punishment, then I feel that you can hardly find any overplays by AlphaGo, even if the humans spend hours for post-game analyses.
The bottom line is that AlphaGo does not operate on an algorithm assuming the less strength of the opponent, as far as my shallow understanding of their paper and their talks go. The algorithm resembles the "healthy" human reading to a great extent in assuming the opponent as strong as herself. If this understanding is correct, then even the opponent AlphaGo cannot punish her own hypothetical overplays severely that easily. That is how AlphaGo heads into dangerous looking moves like the small avalanche on the right top in this video.
That sounds correct to me. AlphaGo's internal meta-game would be shaped just as largely by the games against herself as those against human opponents. Thus, any moves that were too far "overreaching" would be so severely punished by herself that AlphaGo would develop an intuition against the move. There would be a definite limit to the level of risk in other words. It's not like a policy network could train itself into a meta-game where it makes plays in the hopes of a poor response from the opponent. Caution would be necessarily built-in to its intuitions.
Of course, if we further added to the complexity of AlphaGo so that she could remember faces (so to speak), she might well start to make such moves against specific individuals that she knows will allow her to get away with moves she would never otherwise dare to make.