Go Wizardry

All About the Many Aspects of Go
We have millions of friends around the world... and they all play go!

Leela Zero Go-Playing Program


Last week, the night before I was going to my Japanese go club, the South Bay Ki-in, located in the New Gardena Hotel, 1641 W. Redondo Beach Blvd., Gardens, CA 90247; (310) 327-5757, Kevin Carr sent me a text message asking if he should bring Leela Zero to the club. I had no idea what he was talking about.

Kevin explained that this is a go-playing computer program created by the Deep Mind group in London that developed Alpha Go. As the reader probably knows, Alpha Go has beaten all of the top go players of China, Korea and Japan, mostly in matches. However, that program was running on several super computers that cost thousands of dollars an hour to operate. Leela Zero operates on standard personal computers.

On the other hand, Leela Zero is very demanding on the Central Processing Unit (CPU) and memory resources. "The CPU gets so overheated that I have to set up a fan to cool it down," said Kevin. Anyway, I told him that everyone in the club would be interested in seeing it work.

When we got there, Kevin set up his laptop computer (although Kevin is a professional computer programmer, so his laptop is souped up with advanced modifications to the system) while I explained the situation to the members of the club. In just a few minutes it was all set and I invited Tim Chang, a 4~5 dan player, to try it out. Kevin and I then sat down to play other members of the club. From time to time, Kevin would check on the how things were going. Leela Zero has controls that indicate what is happening in the game being played.

Following Tim, we had Mr. Yamaguchi 6 dan play Leela Zero. Then, Mr. Watanabe 7 kyu, played and finally Tony Emsenhuber. None of them were able to beat the program. "Leela Zero has defeated professional players," Kevin pointed out. But the strongest players in our club have also done that! In the Cotsen Open last year, Tony lost to a Chinese professional woman 2 dan player by 1/2 point. That shows that his strength is equivalent to a low dan professional.

Interestingly, Tony also stated that he met the head of the Leela Zero development team at an artificial intelligence (AI) conference. "My function with the Austrian government is to follow the latest developments in technology. AI is one of those that I pay close attention to."

The reader might be interested in whether I played the program. I did not. Kevin invited me to do so, but I wanted all the others in the club to get a chance to play it first. And I did not even follow the games that were played. They hold little interest for me. When I go to the hotel, I always go into the lobby to read the Japanese newspapers. The Meijin title match finished the day before, with Chu U winning the seventh, and last, game. He won the title, which he has not held since a decade ago. That is much more interesting to me. I enjoy seeing the latest innovations played by top players, and to read the analysis in the Japanese newspapers’ go columns.

When Kevin and I discussed the results later, he prefaced things by telling me that he and a friend had tried their luck in playing Leela Zero themselves. "After 30 moves, the Leela Zero controls indicated that the program thought that it had a 90% of winning the game. In Tim’s game, that judgment was given at move 45. In Tony’s game, that point came at move 80, the best that anyone did. And during the game another measurement was given showing that Tony was giving ground in slower increments than the others." "How about Mr. Yamaguchi?" I asked. "He would not let me save the game. I guess he was ashamed of how he played. He said that he knew the mistake he made and let it go at that. Tony said I could save the game, but he asked me not to post it anywhere. I guess he was also ashamed of how he played."

Strange. He had been defeated by a program that had been created by a top AI research group, one with millions of dollars of capital behind it. (Deep Mind was acquired by Google after the Alpha Go matches.) The more games that are available for analysis, the better it will be for devising strategies for defeating these AI programs.

We are bringing Leela Zero to the club again next week. I will have further comments to make here after that. And when someone is brave enough to have the game record revealed publicly, GoWizardry will post it here and I will annotate it.

Now, I continue analyzing the Meijin title match. One other thing. If the reader wonders my qualification for making comments about the game, in the South Bay Ki-in I play as a 5 dan and my record so far in the latest tournament, which began last month, stands at 9 wins and 3 losses. If I win 8 more games, I will be promoted to 6 dan.

43rd Annual Best of Seven Meijin Title Match, Game 6

White: Iyama Yuta, Meijin

Black: Cho U 9 dan

Played on October 21 and 22, 2018.

Komi: 6 1/2 points

195 moves. Black wins by resignation.


Figure 1: Off-Kilter Imitation [Mane] Go (1-30)
Those interested in replaying the game move by move can click here to do so.

White 2 occupies the corner diagonally opposite to Black 1. There is a reason for this. If White played in either of the other corners, Black 2 leads to a complex, fighting game. Cho is known as a strong fighter, so perhaps Iyama wanted to avoid letting Cho get a chance to display his power.

The game proceeds in form of imitation [mane] go that is a little off-kilter. That is, when Black plays 3, White plays 4 in the opposite direction. However, from there on, White continues imitating Black’s moves. It is only at move 22 that White deviates from Black’s lead.

In the moves through White 28, Black gets powerful thickness in the upper left, while forcing White to live on a small scale. One has to conclude that Black is confident of having a solid advantage. Why? Because of Black 29. Playing on the second line in the opening is usually considered bad. And Black 29 is purely defensive; the move does not threaten White in any way. It does not even undermine White’s thickness. Rather, it is designed to prevent White’s thickness becoming any stronger.

White 30 is another territorial move. White takes profit in the belief that Black’s thickness can be neutralized later.


Figure 2: A Complex Game (31-60)

Black 31 through 35 are standard joseki moves, but White 36 exhibits a special strategy. In the standard joseki, White plays on the point that is two points to the right of Black 35. That is because Black’s follow-up is to attach at that point as a forcing move. However, White’s thickness in the lower right means that Black’s attachment or White’s move-in-a-row [narabi] on that point are not urgent moves.

White 36 is played to neutralize Black’s thickness in the upper left. If Black answers by blocking in the upper right at the point above 31, White makes a two point extension to the left of 36, making a base on the upper side.

But Black deals lightly with this by making the capping move of 37 in order to take sente.

In response to the attack on the corner of Black 39, White makes the pincer of 40. Although this is not strictly an essential move, should White make a corner enclosure in the lower left, a Black stone played on the point above White 50 would start to build a large territorial framework [moyo] extending from the upper left into the center of the board. At the same time, Black’s move on the second line in the lower right would prevent White from building a lot of territory on the lower side.

Black attacks the lower left corner a second time with 41. The attachment of White 42 leads to a hybrid attach and extend joseki and attach and block joseki through White 48. Black makes a base in the lower left corner by invading at the 3-3 point with 45, taking even more territory than White did in the upper left.

Next Black moves out into the center with the moves through 57. White’s moves through 58 neutralize Black’s thickness in the upper left to a great extent, but they do not really make any territory. Black has to considered as having the advantage in the game here.


Figure 3: The Value of Thickness (61-90)

It is interesting to see how Black converts the thickness built previously into profit. Before reading the analysis that follows, the reader should examine the board position and see if anything is striking about it.

Here some factors that should be considered: White has territory in four places. (Five if the stones on the upper side are cut off by Black playing atari to the right of 90 and pushes through.) Black also has territory in four places.

Now, roughly count the board. White’s territory on the right side is roughly equivalent to Black’s on the lower side and into the right corner. Black’s and White’s territory on the left side are roughly the same. If that is true, White would be ahead in the game because of the 6 1/2 point komi, right?

Wrong. Black has sente, which in this board position it is equivalent to 10 points. So it is actually Black who is ahead.

However, that is not the most striking thing about this part of the game. What is most remarkable is that Black has kept sente every step of the way. That is the value of thickness. And it is that that is so hard to evaluate. I will always either give my opponent thickness or take it myself because it is so hard to handle it well. Naturally, Cho is a past master of that art.

Black 61 forces White to once again make a small life in a corner. Then Black plays another thick and strong move in the center at 63. This forces White to play at 64. It might seem like White 64 takes a lot of territory, but White’s stones are actually overconcentrated here. White has made barely 15 points. And Black can make the forcing moves of 69 and 71. This is perfect timing to make these moves. White has no choice but to back down with 70 and 72. This proves that the White stones are overconcentrated.

Black then takes more profit in sente with 73 before consolidating the lower side with 75. It may seem that this move gives up sente, but White has to respond with 78 and 80. Otherwise the Black attachment at the point below 78 and one point to the right would be an intolerable forcing move to have to answer.

Black 81 through 89 are a kind of middlegame joseki that build thickness in sente. Notice how well Black’s stones work here in relation to Black 63. Black definitely has the advantage in the center. It will be hard for White to fight in that area.


Figure 4: Give and Take (91-120)

Black plays the forcing moves of 91 and 93 in sente, then invades at 95 on the right side. White’s thin position above makes it impossible to fight back. All that White can do is to Play 96 and 98 to move into the center. But these moves do not threaten Black in any way.

White can only try to fight back with 100 and 102.

Black then hanes at 103 to create some leverage against White’s position. Black has to accept some damage in the lower right corner, but is able to break through on the right side and defend at 111 to make sure that the stones can escape without a problem.

Not only that, but White’s stones in the lower right are now threatened. White has to play defensively to make sure that the stones here live. A sad end for the thickness that White had built up earlier.


Figure 5: Upping the Stakes (121-150)

Black 121 is a move that is very satisfying for Black to play. It consolidates the lower side while forcing White to defend at 122.

Then Black attacks the thinness of White’s position on the right side. Black pushes through at 123 and cuts at 125. White has no choice but to reply with the moves through 128. The result is that White is saddled with weak stones in the center.

Instead of defending in some way, White cuts with 130 and captures at 132. White lost so much territory on the right side, that it is necessary to try to recover in some way.

Then, when Black plays the knight’s move at 133, casting a wide net to ensnare the White stones to the right, White plays loosely at 134. White is ready to sacrifice the stones to the right in exchange for reducing Black’s territory in the center.

Black nonchalantly turns to play the diagonal move at 135, taking territory while threatening to reduce White’s territory above further. It is clear that there is nothing that White can do in the center.

Nonetheless, White ups the stakes on the right with the move at 136. What is the likelihood that this will succeed?

Remember that White’s group in the lower right is not completely settled yet.

Pop quiz: does Black connect at 140 in reply to White 150?


Figure 6: Black Takes More Territory in the Center (151-180) Black 161 takes ko at the marked stone; White 164, same; Black 167, same; White 170, same; Black 173, same

Black simplifies the game with 151 and 153, then hanes at 155 and plays atari at 157. This starts a ko that is dangerous only for White. That is, a flower-viewing ko.

With the moves through Black 175, Black takes even more territory in the center. The invasion of the right side has been a complete success. White’s territory was reduced at the same time as Black’s territory in the center expanded.


Figure 7: Wrapping Things Up (181-195)

Every move by Black increases the advantage, while White can only hope for a miracle. After Black 195, there is no hope left.

195 moves. Black wins by resignation.

Those who wish to comment on the opinions expressed here may send their thoughts to info@GoWizardry.com. The most interesting responses will be addressed in future postings.

Robert J. Terry

Tagged as: , , , , , ,

Leave a Reply


book cover

Go on the Go Collection: Volume I

Three booklets have been assembled into the collection here.

Buy this Book at Amazon

Go For Everyone

Go For Everyone

A New Method for Learning to Play the Game of Go

Buy this book

Book Cover

Journey to the West

This is a semi-autobiographical novel that depicts a unique American success story; a rags to riches tale of a man escaping his humble origins to make millions of dollars, but then he throws it all away due to the ancient character flaw of hubris.

Buy this Book at Amazon