Skip to main content
Loading Explore guides…
AlphaGo from a Go Player · Acedra
Home
Explore
Computer Science
AlphaGo from a Go Player
Sign in to save
Chat
If YouTube asks you to sign in, open the video on YouTube.
Watch on YouTube
Reload player
Transcript
49 segments
Copy
Hide video
0:00
In this video, we're finally going to talk about Alph Go from a Go players perspective. So, I will show you some of the key moments in all Alph Go matches. And even if you've watched the Alph Go documentary, then we will talk about some of the things that weren't covered there. So, here we go. As you probably know, Alph Go was developed by Google's subsidiary, Deep Mind. And because I'm not really tech-savvy, I can only attempt to explain in the simplest of words how that wonderful AI worked. So
0:31
they started by taking a database of human go games and they used it to train Alph Go and they had several neural networks working together on this. The first neural network looked through all the games and it learned to generalize the patterns and predict what the next move was going to be. The prediction wasn't perfect, but it was good enough for now. Another neural network was shown many positions and then it was told who won the game from there. And so it learned to predict who was going to
1:03
win the game from any position that it saw. Again, in both cases, the predictions weren't really accurate. It wasn't 100%. Maybe it was only 60%. Now, armed with this knowledge, the AI would play many, many games against itself, only to gain new data for reinforcement learning. Now that more games were played, the AI would use those games to generalize and find patterns, and then the process would repeat. This is probably a crazy oversimplification on my part. So, if you're an AI developer
1:34
or you just know more about this, you can leave your criticism in the comments. Now, everyone talks about the big match between Alph Go and Lisa Doll, and we'll get to it as well, but let's not forget that there was another match prior to that. So, let's travel back to October 2015. There's a professional go player from China. His name is Funway and he's living in France. And at that time, he became the European Go champion for the third time consecutively. So,
2:01
Deep Mind contacted him and offered to play a match against their new AI. Naturally, he was skeptical, but he agreed to play. His skepticism quickly gave way to surprise, then shock, then despair. He lost five games in a row in that match, 5 to zero. That was astounding. This is the first match lost by a professional goal player to a machine. However, I remember that back then it didn't cause any sensation yet. Instead, this 50-0 result started some whispering and muttering that Funway was
2:32
not maybe a real professional, that he' lived in Europe long enough to forget the skills of an Asian pro. I imagine there must have been a lot of pressure on his shoulders. Go players often say that the AI brought a lot of innovation into go theory in general and opening and joking specifically. However, that Alph Go version was still not quite the ruthless, unbeatable machine that we would know later. This Alph Go played very much like a human here. Let me show
2:59
you a couple of moments from those first games. This is the very first game in the match against Funway. Alph Go is playing white here. This entire opening setup until here looks very human. These jokies on the left were played long before AlphaGo. At this point, it's white's turn and AlphaGo approaches the corner. White responds with a ka and white slides. The most classic of all variations and it's the one that was criticized by AI and it got replaced by this in modern go. However, at that
3:34
time, black maybe exchange here on then protects the corner and Alph Go just plays a normal two space extension. This simple shape that was so incredibly common 10 years ago, you will hardly ever see it now. Now, let's see one more example from game four of their match. In this game, AlphaGo is playing black and look at the joseeki that it chose on the left. Alph Go approaches [Music] attachment. White pulls back. White connects like this. White protects the corner. Black extends. White extends.
4:24
And Alpha Go jumps. And once again, this is something you'll be likely to see in a human game because this jump as protection on the left is not really necessary. There is some invasion here, but it's not really that dangerous. So, in a modern game, you'd be much likely to see a tener key somewhere on the right, for example. And while we're on the topic of this joseki, the modern AI would actually not even connect this way anymore. In the modern game, you're much
4:53
more likely to see a simple solid connection. And after white jumps, black jumps here or here. That is the modern approach. But okay, forget about opening and joking. Alph Go actually made some visible mistakes that Katago today picks up on. Let me show you a couple. Once again, this is game one of their match. Alpha goes white. Fun just invaded at the bottom here. And next, he played here to save these stones. But actually, as we now know, before playing here,
5:24
black could have made a very useful exchange. I could include this example into the useful moves in the second line video. So instead of this, black could push here first. This move doesn't look very powerful, but in fact, it makes the whole white formation weaker and exposes this cutting point. Now, if white blocks like this, which seems like the obvious response, black can actually cut later. And now white wouldn't be able to atari because now black can escape with atari
5:54
and all of these white stones are in danger. Alphago's mistake was that after fun responded here it also failed to see the importance of this move. White should have turned here. This makes the all the white group so much stronger. But instead of this alpha go played here. Black responds. White pushes. White blocks. White saves the stones and black gets a chance to play here. White can't block. So in the game, white had to respond like this. A great exchange for
6:29
black. Now we're still in the same game, but let's skip a few moves forward. Right now, Alph Go plays here to get control of the center. But according to Kato, this is a three-point mistake. And the reason for that is that once again there's an urgent move on the board that both black and white can take. Right now it's black's turn. So punlay could play here which protects the corner and threatens to save the two stones. White would be forced to capture them and
6:58
black and now tenarchy protecting the corner incent. This is amazing. So if we know this instead of this move, white could also play something like this. I know this is a move on the first line and I personally would never be able to find this, but it's capturing these stones and it's threatening the corner. If black doesn't respond right now, this next jump could be quite devastating. And even if black responds, white has moves in the future like this one to enter black's corner
7:30
and reduce it. So, however big this move in the center may seem, that was an urgent point. And by the way, Funway didn't just lose the match peacefully. When he saw he was losing, he tried different strategies. He even used some of the tricks that he learned from Alph Go itself in this very match. But experience teaches us that changing your own style for a particular opponent or event is usually not the best thing to do. And if anything, it makes you play worse. So when we put all five games of
7:58
this match under the microscope of modern AI, we see that the game Punway handled best was the very first one. This game categ's middle game accuracy in this game was higher than that of AlphaGo. He played a great game and his loss was never bigger than 10 points, usually around five or six. For Deep Mind, it was time for new challenges. And next, they decided to challenge the player who had a semi-leendary status at the time, Lisa Doll Non Pro from South Korea, the
8:29
winner of many world championships. Remember all those inaccuracies and humanlike play in this match? Lisa Doll saw the match, too. So, it's no surprise that he couldn't help feeling a little skeptical at that point. He said back then that the level of AlphaGo's first opponent wasn't really on par with his own, and that was the main reason the machine won so easily. In his turn, he was confident about winning the match. The only doubt in his mind was about the
8:55
final score. Would it be 4 to 1 or 5 to zero in his favor? Oh well, DeepMind had some surprises up their sleeve. They had four or five months before the next match to train, train, train Alph Go more. Millions of practice games for the machine. How strong did it become? Nobody outside of Deep Mind knew. And certainly not Lisa Doll. He would only get a chance to discover the current level of Alph Go on March 9th, 2016 on the day of their first game. the first game that he started to play in a
9:25
relaxed manner, not expecting that Alph Go would start a fight in the early opening. Let me remind you how that happened. Alph Go is white here and I remember that the commentators were saying that this looks like a comfortable game for Lisa Doll. Black has a lot of points on the right, these two stones, everything. And the white stones don't seem to be making any points. But AlphaGo had a plan here. White caps. Black tries to get into the center by attaching fully expecting something like
9:55
this where white now has a cutting point and the black stones escape into the center. But Alph Go had a different plan. It peeps here. White has to connect. Next, white pushes and cuts. AlphaGo turns this game into a disconnected running fight right from the beginning. Now Lisa Doll needs to live with those three stones at the top, manage these two stones in the center and also think of something for this stone here on the upper left. At the time, most of the commentators seem to
10:30
be saying that this looks like a difficult fight, but this is a comfortable game for Lisia Doll. And they were saying that right until late in the middle game when it became obvious that Lisa Doll was losing. But if we ask the modern AI about this game, it will show you that actually right from the opening the game was comfortable for white for AlphaGo and that winning gap would only increase after this first fight. However, at the end when it was already clear that Alph
10:55
Go was going to win the first game, something interesting happened. Alph Go seemed to have made a mistake. Let me show you how that happened. At this point in the endame, when everyone was counting that white was a few points ahead, here's what happened. AlphaGo blocked here living with this group and Lucid Doll played the Atari. Now this seemingly starting a co black takes and after this move instead of playing this co Alphago just makes second eye. The commentator said
11:29
that this was probably the first clear mistake that AlphaGo made in this game and it is true. White doesn't have to protect like this. White can keep playing this co considering how many local co- threats white has in this area. But there was something that they didn't know and that was revealed a little later about the style of AlphaGo. AlphaGo's style was a little similar to that of Lie Ho. That is not playing the most aggressive moves killing the opponent but rather trying to aim for
11:58
the most secure path to victory even if it means playing some slack moves and not trying to maximize the win or kill something. So, at this point, these slightly passive moves in the lower left corner only meant that AlphaGo knew it was winning this game, and it didn't really matter if it was winning by 10 or by three points. A win is a win. Lisa Doll was shocked after the first game, and he said at the press conference after it that he believed now the chances were 50/50. Probably at that
12:27
time, he was still underestimating AlphaGo a little bit. Let's take a look at game two. In this game, Alph Go is playing black and the opening looks fairly peaceful until now. Now Lisa doll approaches the corner. Black picks. White stands. White jumps and white wants to make this position a little higher. So white plays here. Having played this move, Lisa Doll left for a little smoke break. And to his surprise, when he came back, he saw this move played by AlphaGo. All the commentators, everyone
13:02
watching this game live and Lisa Doll himself were shocked when they saw this shoulder hit on the fifth line because we all know that a shoulder hit is good when it's leaning on the opponent's position and making it low. So, for example, if white has a stone here, this shoulder hit would be perfectly reasonable, making white low on the third line. But now that white is high, this move seems to be giving white territory on the fourth line. And yet, just like Shuaku's famous ear redne,
13:34
this move seems to be tying the entire black position together. It's helping and developing the lower stones, helping the upper stones, and also preparing to save these stones here and start a fight because at any moment in the future, AlphaGo can connect here and try to attack these white stones separately. This move is preparing for that. Lisa Doll said after the match that it was this move that showed him that AlphaGo could think and play creatively instead of just looking for moves in a big
14:05
database. After losing the second game, Lisa Doll was put on the verge of losing the entire match. So, game three was supposed to be his last stand and no surprise there, he started a very big fight in the opening. AlphaGo is white in this game and it just played here. Lisa Doll responds like this and white responds with a very light move to jump out of here. And would you just look at that? Lisa doll attaches trying to separate the white stones and start a fight. AlphaGo parries this way. Black
14:42
separates. White pushes. Black saves the corner. The black and white stones are all disconnected and they will keep on fighting like this for the next 50 moves or so. And this is the result of that fight that AlphaGo handled beautifully. Nobody died in the process, but AlphaGo got a big moyo at the bottom and a comfortable lead in this game. Now we skip a few moves forward and this is the same game closer to the endame. Alpha goes white and white has too many points
15:10
at the bottom. At this point, Lisa Doll feels that if he doesn't invade this area, if he doesn't live here, it wouldn't be enough territory for him. So, he makes this exchange first. AlphaGo answers. Then a few more exchanges in other
15:37
places. Now this atari first white ignores and saves this stone. And now the final fight, an invasion. If Lisa doll lives here, all of white's points will be gone and AlphaGo will lose this game. So, white needs to kill these stones. Let's see how that's going to work. White limits the space. Black attaches. White separates these stones. Black attaches here. There's some weaknesses. So, white connects. Now, a tricky move to the second line. Preparing to make shape to the top and to play
16:24
here. Alph Go
16:33
separates. Leafy doll makes a double tiger's mouth. White needs to block. Black pushes. This is white's sente. And white connects. Now [Music] push and cut. And these white stones are also separated from the rest of the groups. White blocks. Black plays this honey. White responds and black makes an eye and prepares for a co. Many people believed at the time that machines were weak at handling co-fights because you know a co involves a lot of factors. You need to calculate
17:24
all of the possible co- threats and evaluate which one to respond to and which one not to respond to. So we can almost say that this co-ight was the final test for AlphaGo and it did just fine. After this co-fight, Lisa Doll resigned the game. The score in the match became three to zero and AlphaGo won the match. But believe it or not, and I think many of you know what I'm talking about, the most spectacular move in the match was yet to be played in game four. In this game, AlphaGo is
17:53
playing black. And black has a scary territory/mo at the top of the board. Lacy Doll did a very good job at taking a lot of points everywhere and leaving those four stones up there lightly, hoping to save them later in the game. And the moment to show some miracles of Sabaki is now AlphaGo just played here trying to just swallow those stones in there. White plays here threatening to rescue them and AlphaGo makes a very solid shape. Connecting here would be impossible for white. Black is too
18:29
strong around. So, white cuts here as a probe and black responds this way. Just separating all the white stones. Another useful exchange. Black blocks this. Black captures that stone. And next comes the move that only led Doll could see. white wedges. This move is like a bomb explosion. It just activates all the AGI in this shape at the same time. Of course, now that we have superhuman AI to tell us how things could have been played, we know that actually black is
19:12
leading by about 10 points here and nothing is changed by this amazing tuji. Black should respond like this and after this white shouldn't be able to get any advantage out of this. But in that particular game, at that moment, that version of Alph Go couldn't find the right move. It played here. And this allows white to play
19:38
this and then cut. Now, there's a possible throw in and snap back here, a cut here, this atarian cut here, this atari. All of those possibilities allow white to escape with these stones through here. Save everything. Alph Go started making unreasonable moves, not being able to understand how to win the game anymore. And Lisa Doll won this game in the end. Lisa Doll said after this game that he had won many world championships, but he was never congratulated and greeted this
20:12
much after winning just one game. [Music]
20:22
[Applause] [Music] [Applause] In the end, Alph Go won this match 4 to1 and Deep Mind showed the whole world that the AI was now stronger in Go than human players. But Deep Mind was not going to stop there. The alpha go that played against Lisa Doll was stronger than the one that played against Fon Huay. So what if we give it more time and let it train more? In December that same year 2016, we got an answer to this question. A mysterious player registered on Taien Go server and played 60 games
20:59
against all the strongest players in the world and won all 60. Of course, this was another version of Alph Go that became known as Alph Go Master. Another match between that version of Alph Go and China's strongest player Ko was held the following year and Ko never had a single chance. That match was 3 to zero. Alph Go got to superhuman level. Did Deep Mind stop there? No, they didn't. They released yet another version of Alph Go in October the same year, 2017.
21:31
It was called Alph Go Zero and it was entirely different from all previous versions because no human games were used for its training. It didn't start from any human games at all. It developed all of its skills and strategies from scratch. So they it started playing with itself just making random moves at first and then little by little it surpassed all previous versions and then got above them. It played a match against the previous strongest version Alph Go Master. out of
22:01
100 games and it won 89 of them and lost only 11. This shows that maybe we're somewhat lucky that that Alph Go version never played against any human player. The results that it showed indicate that maybe it was about three stones stronger than Alph Go Lee, for example, which means that if Deep Mind had waited another year and allowed Alph Go Zero to play that match against Lisa Doll, that match wouldn't have been that pretty and it wouldn't have been very close either.
22:32
Once DeepMind managed this gargantuan task of developing a superhuman Go AI, they published their findings and that allowed other developers to replicate their process and make equally strong AI. Thanks to that, we now have several of them. The strongest ones currently being fine art and galaxy from China. But at that level, it doesn't really matter which one is the strongest. Even Kate that is easily available for free plays at the level that exceeds that of
22:58
any human professional player. Lisa Doll said something interesting in one of his interviews after the big match. Something that struck a sad chord in my Go player heart. People always believed, he said, that they were creative and Go was the perfect playground to show their creativity. But Alph Go demonstrated that people were actually very conservative and conventional in their playing and in their judgment of go positions. From teacher to student, knowledge would be passed on. Most of us
23:27
would just remember that this is how you need to play here. The funny thing is that now with AI being the new benchmark of perfection, we've become just as conventional except with a different set of moves. Alph Go made a revolution in the Go world. It changed opening theory forever. A lot of new variations and ideas were introduced that were surprising in 2017, but now less than 10 years later, all of them are accepted as the new golden standard. The changes
23:54
that the AI brought into the Go world are so monumental and so numerous that we'll have to take a very good look at them in some of our future videos. Until then, keep making creative magic on the Goboard and don't overdo it with those sunsign invasions. Okay. By the way, you can also watch these lessons on our platform gomagic.org except there you'll watch them with interactive quizzes right within the lessons and practical exercises right after them. And if you
24:22
enjoy watching these Go videos and you don't want to miss others like this one, go smash that like button, subscribe to our YouTube channel, and this is Go Magic. [Music]
24:42
[Music]
Here to help you learn
Sign in to keep a private conversation with your course tutor.
Ask me anything about the material…
Sign in to chat