AlphaGo vs. AlphaGo with Michael Redmond 9p: Game 1
Vložit
- čas přidán 30. 07. 2017
- Michael Redmond 9p, hosted by the AGA E-Journal's Chris Garlock, review the 1st game of the amazing AlphaGo vs. AlphaGo selfplay games. Fully commented sgf file -- with lots of variations -- is posted at www.usgo.org/news/2017/08/redm...
The 50 game series was published by Deepmind after AlphaGo's victory over world champion Ke Jie 9p in May 2017.
Produced by Michael Wanek & Andrew Jackson
Thumbnail image of Rock 'em Sock 'em Robots by Lorie Shaull - Own work, CC BY-SA 4.0, commons.wikimedia.org/w/index... - Hry
I'm really excited about this new series.
As a kyu player, this looks like a game from another planet. Chris's face at 52:14 is a picture of how I felt throughout the whole game. But I love it!
Thank you Michael for the hard work.
As a near-pro level amateur, this looks like a game from another planet too!!!
I love that you mentioned Cho Chikun for the real reasons behind his moves (in his youth strictly speaking) that look territory oriented and thin for many.
WOW that peep that he explained towards the end, of the danger it held, and how W responded correctly... was unbelievable.
Fantastic intro and a most fantastic review by one of my favourite players! Thank you so much!
Amazing. This is a great gift for the community you're doing, it's like listening to a lecture about a classical symphony to help get some idea about what is going on at various levels.
First of all, I want to say thank you both for the content you have provided that I wish existed many years before. I really appreciate the amount of effort Michael Redmond has put into these games and in teaching go in general. The variations shown and explaining/simplifying the moves is done so well. I'm glad that I found this channel and hope that you will keep providing great content.
This is much appreciated. Thank you Michael & Chris.
Thank you very much for your effort in these reviews Mr Redmond, it is most appreciated. As an upper kyuu player I find watching these reviews very rewarding.
Thanks for the deep analysis Mr. Redmond! And great hosting Chris!
Thank you so much for a wonderful review! I love that you didn't limit the time and played out the endgame.
So generous to share this, and a delight to watch! One a week would already be a treat, so take it easy if you need to...
you both make a big team. I'm more and more a fan of Redmond's reviews, so clear despite the complications.... thank you a lot!
Very good review, you two have really improved your commentaries since the beginning. Keep it up !
That is a great intro! Seriously impressive!
Really really valuable, both entertaining and enlightening, to hear Michael's perspective on this game, which I've seen a lot of commentary on. I feel like Michael has wonderful insights into where alphago is at.
Awesome!! Thank you very much for the review.
Absolutely wonderful commentary Mr. Redmond!
I hope that in not too distant future, programs with the strength of AlphaGo, or maybe even the learned knowledge of AlphaGo itself, will be available for purchase. Then professionals like Mr. Redmond would be able to much more efficiently analyze games, either between AlphaGo, or human games.
That is the way it happened in chess. Top grandmasters use programs to help them analyze.
He did mention the possibility of Deep Mind releasing the value net data for these games which would definitely help his analysis.
I actually wrote the lead programmer of AlphaGo, David Silver, about this, and he honored me with a response, that yes, they will make the learned knowledge available in some way.
Hi person from the future here 👋 (5 years into the future actually), this has actually happened! In newer videos Michael uses programs like KataGo to help him analyze these AlphaGo vs AlphaGo games!
yeah it's finally here! and love the intro!
I think it is very difficult to understand, but let's say I get at least 10% of what you are trying to impart here, and that is still very fun. I like that although the content is serious, these guys still have alot of fun. I'm excited to watch them all. Thank you for the hard work!
I hope that your classes will be available online, Michael Redmond! Wish I could attend.
Plot twist: Those 50 games wasn't played by alpha go but at the local go club.
Plot twist: grass is poisonous
Thank you so much Mr Redmond and AGA. Can we expect more games after these four?? (Really wonderful commentary).
Didn't start watching yet, but super excited!
Thanks for this
woo that intro tho
iamstickymouse Splines have been reticulated.
Definitely converged on truth!
no wonder loading joseki gave error
It's fascinating how MR doesnt even comment on the five diagonal stones because it makes such perfect sense to him in the context of the game, whereas Mr. Sibicky just about lost his marbels. Love the analysis even if it's so immensely difficult to understand all the intricacies of the game. And just the contrast with the excellent and equally lovely N. Sibicky, I think, is highly informative.
Yes! Finally!
Hi Chris great interview
good to c u again
your Canadian friend
Marty
I love this review, but I'm looking forward far more to reviews of the other 49 games as this one was reviewed by several people.
It'd be hilarious if both sides of Alpha Go passed before even placing any stone on the board. It'd be like "Screw this, I can't outplay myself, I'm just too darn good."
It seems that the most logical understanding of your misunderstandings, shown in your "humorous" perspective of playing against oneself, in order to find the faults in our choices/thinking, in order to have better choices/thoughts. Is what has lead you to such an odd understanding of playing against oneself or a program play against itself to get benefits/ gaining inability. Not to waste time, since the program is learning to do better and better. It still doesn't understand everything about this game after all this time according to the information we have at this moment. That's why we keep, and it keeps beating itself, instead of coming to a draw every time. Though you should know there is a tiebreaker for going second since there is an advantage to going first that will allow the person going first or the program going first, to always win, if there wasn't at least a half a point tiebreaker, when completely evenly-matched.
Wow, that's really well-made intro
All the nerds in the audience (myself included) are losing their minds for that awesome intro :) Thumbs up if you've computed a Zobrist hash!
Chris where did you get 220k TPUs my man? You gonna share those with the rest of us?
after the exchanges roundabouts 43:00, what do you think about the value of sente for white and the way it was used in the upper right? was it worth the exchange of 5 stones in the center and the relatively low point value of the living group on the left?
I'm happy and satisfied that this Cho Chikun/AlphaGo interpretation of thickness is so straightforward and gets rid of the fluffy conversations that seem to focus on what's not a good explanation of thickness.
I missed the point where the game became favorable for White.
I would expect the AI's win percentages to remain extremely close to 50% throughout, since any deviation from that would imply that one side had found an effective move which the other side didn't see - but with identical AIs, that seems extremely unlikely. (Though I'm unsure how much role Monte Carlo still plays in this version.)
It's a great pity that DeepMind is retiring AlphaGo; there are so many wonderful experiments I'd love to see, such as letting multiple versions self-play independently without ever playing each other and seeing if they develop distinct styles (and if one somehow ends up stronger).
Does anyone know what joseki Michael Redmond is referring to shortly after 20:00? One that he feels is worth learning. These are great, thank you for all the work being put into this.
I think he said tsuke-hiki, roughly translating to contact-draw back. I believe it refers to the shape where for instance when one makes a high approach to a 3-4 point, to contact underneath and then draw back.
The Dream Team
if only I could 'like' a 100,000 times, it would not be enough.
At 31 minutes, Michael mentions that he made this into an sgf file, but is it available for download somewhere?
Glad to hear that Michael is enjoying the review process so much, due to the challenge posed by such unfamiliar play. This promises to be quite fun and fascinating to watch. One thing I have to comment on is the (over)use of the term "simplify". I don't feel that this description actually applies at all to AlphaGo, or even to top professionals in Go. I think that it would be much more correct to say that the player in the lead in a top-level Go game is "limiting" or perhaps even "smothering" the opponent, by choosing the moves that leaves their adversary with fewer options to find winning moves/variations. This strategy is called "prophylaxis" in chess, and is fairly well known and studied.
Simply put: the idea that an AI capable of playing an entire Go game against itself in three seconds would need to simplify the game to win just sounds silly. I can understand that the descriptions ("limiting" vs "simplifying") may amount to the same thing in some people's minds, but I think that using "simplify" to describe AlphaGo leads to a very false impression of its actual strategy.
I would be one of those to say that "limiting" the number of winning variations is, for all intents and purposes, the same thing as "simplifying". If we're going to argue terms that apply or do not apply to AlphaGo, I would say that saying AlphaGo has any "strategy" whatsoever is misleading, as it's just playing what moves the value network rates most highly. But of course, for a human's purposes, that's virtually the same thing in the same way that I would say "limiting" and "simplifying" are the same thing.
I disagree that anthropomorphizing AlphaGo is misleading. It is only misleading if it is done poorly, which in the case of saying that AlphaGo "simplifies" the game is a poor description of what it has accomplished. Of course it has a winning strategy...it wouldn't be so successful unless it had found the most efficient and certain path to victory. I would propose that this strategy is to take a lead and then smother/limit the opponent. That is the essential purpose of prophylaxis in chess. A pawn lead is almost inconsequential at the 10th move, but after almost all the other pieces have been removed from the board, that pawn lead is enough to win the game.
I'm not necessarily arguing that it's misleading. My argument is that it's only "misleading" in the same way that "simplifying" is "misleading", which I argue is not.
Now, if we take a look at the raw AI on the other hand, it strictly speaking has no strategy, because it's only trying to maximize winning percentages. Saying it has a strategy would imply that it's going through logical steps (e.g. "I'm trying to make life here to limit B's influence to the center, so I will play here"), it's just we describe it like that because that's how our brains work and is what we can understand.
I'm assuming that you come from a chess background, and saying "simplifying" doesn't really fit what a chess player would call it. But that's simply (no pun intended) the way Go players describe it.
Alphago is one step. Will Alphago explaining why he chose a move be the last step? I hope we reach that one in my lifetime.
With all these super close scores, including one game against Kie Jie, I wonder if Elwyn Berlekamp's "chilling gets the last point" methodology is applicable in any of these games.
Does anyone know the name of joseki Michael Redmond recommends learning at 20:06? Thanks in advance
I think he's saying "the tsuke-hiki" joseki, which is the 3-4 point high approach followed by attach 5-3 hane 6-3 draw back 4-3 joseki.
thanks to michael sensei and chris garlic, sorry i allways wanted to say that :) you are cool too
Of course the game is balanced when both players are exactly equally strong
Amateurs of the equal rank will have scores that vary a lot more. It's only for top players in the world that half a point really makes a change in outcome.
it seems like Alphago is fighting over miai points (at a large scale) throughout the whole game
Sometimes it plays slow-looking moves that create miai so that it can grab the last large point, or moves that alter the value of the last two big points.
go aint a flip of a coin game.... it is pure mathemathical....
i now wonder, what is youre actual point? or you just wanna feel smart? the last thing i can give you. oh wow such smart comment you made.... now you have to explain youre point
No, for God, go would just be another version of tic-tac-toe. The outcome is predetermined and either black or white would win every time (considering the use of half point komi).
loading joseki ... error... I like that line
LOL the intro
Who the hell made that intro?
Edit: And who the hell votes this down?
I did it :)
Then you are talented Andrew! Unless you meant "I am the one voting it down".
It was probably Dwyrin voting it down for AG weirdness. :D
At 40:35 does Redmond just mean A and B are miai to win the capture race? Because if black gets B first it looks like the ko makes a false eye.
If Black takes B, White takes A, Black takes Ko, then White cuts and captures the two black stones towards the center to make an eye.
if Black connects the two stones, White fills in Ko to make two eyes.
W cannot cut or capture those two black stones toward the center even if black takes ko. That would take two moves.
Oops, I misread there. Didn't notice the one white stone was in atari.
It seems you may be right.
There are two eyes in both cases. There is always an eye at C16. Black plays B, white makes the eye at D15. Black plays A, white makes the eye at E16.
If black plays B, there is no eye at D15 because B will start the ko to make it a false eye.
i dont even get the rules yet but some times i click a go vid
Try playgo.to/iwtg/en/
This is nightmare food
I literally do not understand anything whatsoever
new at go?
USGO.org. Is a good place to start learning this game. You can also usually find people to help you learn the game on Kiseido.com. / GOKGS.com
#badajiinthecorner
We are still waiting for a legit human vs IA match ... I don't know what the people from Deep Mind are so afraid of ...
Mushmurok Zangief IA? If you mean the AI AlphaGo, there have been 2
And there were 60 fast online games, brutal
this same video is been released 2 times earlier.... i mean ive seen this review before
54:25 LeelaZero thinks R17 is a mistake and prefers P16 (black loses the lead here)
Can Michael beat Alphaho these days?
Can Michael beat Lee Sedol or Ke Jie?
Thofte100 Yes
S L AlphaGo has far surpassed any human.
I think he is wrong about the reasons alphago does not play josekis
My guess is that since Alphago's self play games are 5s per move, that even it cannot read out complicated joseki in that short amount of time, so it perhaps had poorer results, relatively speaking. But then again who knows, it obviously plays very complicated middle games, so that guess doesn't really make sense.
Alpha Go lost! omg