Before Game 5

translated from original article in Chinese:写在第五局之前

— Li, Zhe (6p)

(translated by: Chun Sun 5d, Michael Chen 7d, Yi Tong 3d)

I would like to start with reviewing my previous posts. Then I will present my analysis about Lee Sedol’s strategies and actions in Game 3 and Game 4. Finally, I want to share my personal perspective on this man-vs-machine series as a professional player.

0. Review of my previous posts.

In my first post “Lee Sedol’s Strategy and AlphaGo’s weakness”, We analyzed Lee’s probe and strategies when he was matched against AlphaGo for the very first time, and explored AlphaGo’s mistakes and good moves in a “traditional” sense. Then we discussed the reason behind the “mistakes” from the algorithm’s perspective, and revealed the difference in definitions of “mistakes” between human and computer. We pointed out two weaknesses of AlphaGo, based on the game analysis: 1) Lack of logic; 2) Tendency to avoid ko fights. We raised our suggestions for Lee for the second game: Routine opening, and actively look for kos.

In my second post “Nobody could have done a better job than Lee Sedol”, I started off by arguing why it’s impossible to have a so-called “contract of prohibiting ko”.

Then we discussed that the AlphaGo plays based on its judgement of “win rate”, rather than “best points” with samples in the first game. We also explained Lee’s plan in the first game. Based on that, we analyzed Lee’s strategies in the second Game. We pointed out AlphaGo’s true strength (and its imperfection), and the reason why its strength was underestimated by us humans before the series. We showed Lee Sedol’s efforts in the first two games. We also raised the importance of having Go AI as reference or companion for human to explore the game of Go. Finally, we predicted the possible outcomes of the final three games, and how we should and should not learn from AlphaGo.

1. Lee Sedol’s strategy in game 3.

Let’s start. In this game, we imagine ourselves into Lee’s position, and try to experience what he had thought and did during game 3 and game 4, and how he managed to get a win.

During the first two games, Lee used different strategies to obtain important information, we summarize these as below:

Game 1:

1.  From using an uncommon opening, he knew that AlphaGo doesn’t merely memorize past game records.
2. From playing away from joseki, he knew that AlphaGo plays with global consideration.
3. From leading the game into open-and-complex situation, he knew that AlphaGo plays well in complex battles.
4. From playing a close game, he knew that AlphaGo retreats and loses points when it thinks it’s leading (because of the lack of logic).

Game 2:

1. From using routine opening, he knew that AlphaGo may play creatively.
2. From playing a very balanced game, he knew that AlphaGo is extremely strong in the second half.
3. From some local exchanges, he knew that AlphaGo avoids ko fight. (It could be shown in Game 1 too, but it was more obvious in Game 2)

If you had obtained this information, how would you play game 3?

I predicted that Lee would try ko fights, although it might be useless. I also thought opening from the center would be interesting, as we might see something new.

However, to get a win, routine opening is still more promising than uncommon ones. This is because routine openings are derived from the collective wisdom of human players. Since AlphaGo doesn’t play normal openings anyway, choosing uncommon ones won’t have any advantage.

So we have decided what opening to use. Then what would be the follow-up strategy? We know that after a balanced position in the middle game, AlphaGo plays really well in the second half without “mistakes”. It would be difficult to win if we choose a balanced game. So we need to hope for a big advantage before the middle of the game.

Given 1) normal opening, 2) shooting for a quick lead, a common recipe would be: the cosmos style!

Lee chose high Chinese with an exchange of corner approach in the lower left, this is the most typical cosmos style nowadays.

The key characteristics of the cosmos style are: 1) Easy to get into a bloody fighting game. 2) The difference between a success or fail execution is quite large.

As expected, AlphaGo invaded at move 12, and gave black a chance to engage in a battle in the very beginning.

Lee’s move 13, 15 were the most aggressive play among all options.

After second game, Lee was questioned about the absence of his usual aggressiveness in the game, about his way of “losing peacefully”. In my opinion, it was only Lee’s strategy. He found some possible weaknesses of the machine in game 1, and attacked these in game 2. He carried out the plan thoroughly even though the plan was not of his favorite style (The master of balanced close game is Lee Chang-ho). That being said, specific strategy did restrain him, together with the inevitable psychological pressure, so he couldn’t play at his peak level. And that was what he should not be blamed for.

Another voice was, why do we look for weaknesses of a computer? He could just have played his own game. It’s already losing if you try to look for your opponents’ weakness in the beginning.

Every player will face difficult choices during the game. No matter what you choose, you are being yourself. To quote from <The Art of War>:”If you know the enemy and know yourself, you will never be defeated”. To understand your opponent is not only for the win, but also for the respect that your opponent deserves.

Without Lee Sedol’s showcase of different strategies in the first three games, we could not have known AlphaGo’s characteristics on the game as complete as we can today. He could not have won the fourth game with agony but joy. Think about this, what kind of misunderstanding we would have towards AlphaGo if we saw game 4 in the first match?

Let’s go back to the game 3. Lee opened with cosmos style, attacked in the most aggressive way when AlphaGo invaded, with the hope to gain an early lead.

However, AlphaGo had a near-perfect counter attack.

White’s move 16 was a tremendous counter-play to black’s 15. In the following sequence, move 21 was questionable. However, it’s something you only get to know after the sequence gets fully played out.

White’s move 32 declared the failure of Lee’s strategy in this game.

For AlphaGo, it doesn’t possess the concept of “to make life for an orphan group”. However, in game 3, it showed its ability in this aspect when confronted with Lee Sedol’s attack.

Almost all professional players admitted that AlphaGo’s strength is above human players after game 3. This was because AlphaGo crushed Lee Sedol with an unquestionable victory in game 3, in a way that human players could understand.

After white’s successful counterattack in game 3, Lee’s black groups were already broken. However, Lee showed his strength: He made new probes even under this circumstance.

Black’s move 77 was a test of AlphaGo’s ability to understand the interaction between the local and the global.

Black’s move 115 and onward was a test of AlphaGo’s strength to kill a group. Sometimes Lee Sedol plays like this against human to put on some test when he’s behind, this time, the move was targeted on the machine.

Finally, Lee was able to make a ko in the lower part. Although the result was related to AlphaGo’s conservative moves, it was something that only Lee Sedol could pull off.

If we already saw a great first half of AlphaGo, we may not be able to see a similarly great second half in the same game. As a player, I might feel it’s regrettable, but it’s based on the algorithms.

Maybe for someone, the ko fights later in this game disproved of the claim that AlphaGo can’t fight a ko, or that there was a contract to forbid ko fights. But for me, it was just showing a fact that can be deduced with logic.

From information he gained from the first two games, Lee chose a strategy to try to gain some lead in the first half by using cosmos style. Lee was a little helpless for having chosen this strategy, because after the second game, he lamented that he “can’t find it’s weakness”. So he had to deduce from the fact that “AlphaGo played fantastic second half”, and concluded that he had to take lead in the first half.

However, the loss in the third game was not completely useless, it set the stage for Lee’s win in the fourth game.

2. Lee Sedol’s Strategy in Game 4

Overall, Lee Sedol tested AlphaGo on three aspects: fighting, close game, taking care of groups in danger. AlphaGo did well in all three.

In the post-game press conference after game 3, Lee didn’t sound so clueless like he was after game 2. He found perhaps the last strategy against AlphaGo: To let it kill.

That is to say, Lee would lead to such a situation that he would defend an orphan group and seek for life.

This is one of the only options left, the retreat of AlphaGo in game 3 in the lower part gave some confidence to Lee.

In terms of calculation needed, offense is more difficult than defense, because to make life, you only need to calculate for your own group and find one single variation that works. On the contrary, to kill a group, you need to consider all the possible variations that your opponent may take to make life. For AlphaGo, the search space increases.

It was very clever of Lee Sedol to choose this strategy.

It was indeed the move that AlphaGo didn’t consider (very low priority in the search ranking), the move 78, that triggered AlphaGo’s collapse and turned the game over.

---

AlphaGo’s move 23, attach, was an unprecedented move for professional players. I’ll explain in later posts about why it’s so in more details.

Anyway, Lee chose territory in the left, and left all the thickness to black. White’s moves were too conservative according to professional players’ standard. However, considering Lee’s strategy, and some understanding to that move 23, we would know later why Lee played like that.

White’s move 40 started the invasion. This was a typical “gain territory then invade” strategy. It forced AlphaGo to attack and kill.

Black’s move 47-51 were something interesting and worth learning. Again I’m leaving the technical discussions to the next post.

Then, “The divine move”, move 78!

A lot of professional players have explained why this move doesn’t work. Black had many ways to defend and to maintain the lead.

For example, atari from position 1, white couldn’t make life.

Another way was to atari from below at position 1, in this case, white could connect with a ko, but after black’s move 21, which leads to a tradeoff, black would have won the game too.

Move 78 wasn’t a move that works, it was a move that shines.

This was a moment where human intelligence burst out.

This move very likely satisfied the following conditions:

1. Moving away from the search space that AlphaGo had calculated previously.
2. Increasing possibilities, making more branches.
3. Involving possible ko.

It’s very difficult to find a move that satisfied all 3 conditions.

As for why AlphaGo failed during the game, please read the post-game analysis by Dr. Yuandong Tian (Facebook AI Research, author of computer Go program “darkforest”), or wait for Deepmind’s official response after the series is over.

In the previous post, I said the following about Lee:

“Matched against AlphaGo, human champion Lee Sedol didn’t have even the slightest contempt, he prepared well. He got rid of the arrogance and bias that any other would have, he tried to understand the mechanism of AlphaGo, and to find its weaknesses. From the beginning, he attacked directly into AlphaGo’s possible weaknesses, and he adjusted quickly after failure. He soon started his second, and third well targeted attacks. It is because of his strategy, humans saw AlphaGo’s strength, style, and decision making that is very enlightening. In the second game, he already found a strategy to get close or even surpass AlphaGo in the middle game, thus, for the very first time, we were able to see the spectacular and dreamy second half strength of AlphaGo.”

After game 4, I believe more people will understand my comment.

I’d like to make a new comment for Lee about his game 3 and game 4:

“Lee Sedol kept seeking strategies to attack AlphaGo when there was no weakness to be found. After three losses in three different ways, Lee tried a new strategy in game 4 and executed it well. This time, he succeeded. When AlphaGo saw the ghostly move 78, it exposed a weakness that was possible to be exploited. We are still not sure under which condition this weakness will be exposed, but Lee Sedol succeeded this time. The winning of game 4, was the best reward for his endeavors in the first three games.”

3. As a professional player, how do I look at this man-vs-machine series.

For me, the most astounding part comes from the game records. AlphaGo played a lot of moves outside of human experience, moves that weren’t understood by many professional players.

However, I start to see a spectacular world beyond these moves, a world that was not created by AlphaGo; a world that has been hiding inside of the game of Go itself for thousands of years.

AlphaGo didn’t solve the mystery of Go, in fact, it is far from doing that. AlphaGo only revealed a new world to us. How long do we need to fully understand this world? A year? Five years? Or a decade? I believe that humans will be able to understand these data from AlphaGo with human logic and rational thinking. In fact, the birth of AlphaGo is the exact proof of the importance of human rationality.

Go board is a closed space. The number of variations is astronomical, but still finite. As I said in the previous post, if the exhaustive search of all possibilities is “sky”, and learning just the rules is “ground”, where do we stand? It’s a question to anyone who loves Go. The birth of AlphaGo gave us a reference point. It made us not alone between sky and ground. It made us one-step closer to the eternal truth.

This was the very reason I burst into happy tears and rejoiced after game 2.

Thank you Lee Sedol. Thank you AlphaGo.

At the end of my first post, I concluded:

If we keep using the narrow and constrictive part of established human understanding to evaluate AlphaGo, we will never even figure out how we lost.

Now more and more players know how AlphaGo plays this game, what we need to do as the next step is:

“Let’s understand the new world that AlphaGo just showed us with our reasoning and logic. This is the most sincere respect from a professional player to the Game of Go and to human rationality.”

Looking forward to the historic game 5!