Pages

Sunday, May 1, 2016

Part 10: Review of Game 5: AlphaGo unfamiliar with common tesuji in ultimate moyo game (The historic match of deep learning AlphaGo vs. Lee Sedol)


[NL Versie]

Review of Game 5:  AlphaGo unfamiliar with common tesuji in ultimate moyo game  


Professional  Kwon Gap-yong (8p), who also has been a mentor to Lee Sedol, said before the game: "Lee has already learned that AlphaGo tends to make strange moves when it faces unfavorable conditions. Lee will have more of a chance of winning the final match if he makes key moves that are difficult to read".


In the fourth game, Lee Sedol exposed some of AlphaGo's weaknesses in judging the amount of aji in the program's center moyo. When analyzing the game afterwards with his team, in the exhilaration of victory, Lee Sedol must have concluded that AlphaGo's moyo treatment had been far from optimal and could be put through the mill once again this last game of the match.   


AlphaGo starts off with a probing move (yosu-miru) to see how Lee Sedol will answer (circle in Dia. 1). On that basis white will determine if and how to complete the joseki in the bottom right. 


Dia. 1:  Game 5, after white 12 (circle, Lee Sedol is black) 
Black replies calmly, then plays tenuki to prevent white from making a double extension to the right (circle in Dia. 2). Lee Sedol presents AlphaGo with a choice: clamp and allow black to attack the three white stones in the bottom right, or try to fight and prevent black from connecting underneath?


Dia. 2:  Game 5, after black 17 (circle, Lee Sedol is black) 
AlphaGo chooses to clamp and makes a base for it's three stones in the upper right (see Dia. 3). Then both sides cut and AlphaGo ends in sente as Lee Sedol connects his attacking stones underneath (circle in Dia. 3).


Dia. 3:  Game 5, after black 25 (circle, Lee Sedol is black) 
Black gets a rather large and secure corner in exchange for white's strong and balanced group at the right edge. After moves on the left, where Lee Sedol builds a solid group and white makes a well-balanced extension from the bottom left corner (Dia. 4), AlphaGo plays to make influence towards the center and to build potential in the bottom left quarter of the board (circle in Dia. 4).

At this point, it is completely unclear how many points white's move is worth (how much it could add up to) but in any case this move works well with white's other stones on the board.  And pushes Lee Sedol to prevent that AlphaGo will play nobi or keima downwards to further build it's moyo. 

Both Lee Sedol and AlphaGo play a very solid opening and after move 40 (circle in Dia. 4) the outcome of the game appears completely open from their opposing moyo and territory strategies.


Dia. 4:  Game 5, after white 40 (circle, Lee Sedol is black) 
Lee Sedol avoids that AlphaGo will foreclose the bottom right corner and extends his corner (Dia. 5). AlphaGo calculates that it's three stones have enough aji to make a fight meaningful. The program apparently is unfamiliar with a  common tesuji (also known as the 'tombstone squeeze': you offer two stones, subsequently throw in another one, in order to rob the opponent efficiently from the inner liberties).

This tesuji frequently occurs in the games AlphaGo originally has been trained on, so it is quite incredible that the program didn't learn to handle this tesuji correctly. AlphaGo looses points here (as well as ko threats), apart from the unnecessary waste of options to utilize the potential aji  of it's three white stones (Dia. 5).

At this point in the game, Demis Hassabis tweeted: "AlphaGo made a bad mistake early in the game (it didn't know a known tesuji) but now it is trying hard to claw it back... nail-biting". He added later: "The tesuji itself appears quite often in pro game records, thus it is most likely learned through the 'policy' network. I personally feel AlphaGo should be aware of the tesuji itself, but Lee Sedol's follow-up move was probably better than expected by AlphaGo". He bites his nails and hopes that AlphaGo very soon gets a chance to make things right again (Dia. 5). 


Dia. 5:  Game 5, after black 59 (circle, Lee Sedol is black) 
Fortunately for AlphaGo, it played these moves in territory already more or less realized by Lee Sedol, and it has a few sente moves on the outside that might come in handy later if white wants to make points there. 


Dia. 6:  Game 5, after white 70 (circle, Lee Sedol is black) 
White especially focuses on influence and, after a few moves in the upper right corner, a battle develops for AlphaGo's meanwhile firm center moyo (Dia. 6). 

Lee Sedol invades to pressurize AlphaGo further in the upper left corner and to reduce AlphaGo's potential center moyo at the same time (triangle in Dia. 6). However, AlphaGo plays a beautiful and effective response which turns around the flow of the game immediately. Now, Lee Sedol is put himself under very high pressure (circle in Dia. 6).

AlphaGo's move prevents black's invading stone to connect easily with black's group to the left, prevents an escape of the invading stone to the center to make eventually a base there, forces the invading stone towards white's strength top right, and firmly contributes to it's built up sphere of influence in the center. Moreover, the idea is that if Lee Sedol tries to live, AlphaGo can exert huge pressure on him to become strong on the outside (to reinforce it's moyo even further, see Dia. 6). 


Dia. 7:  Game 5, after black 81 (circle, Lee Sedol is black) 
There's no other choice for Lee Sedol than to create a base. But in exchange AlphaGo gradually continues the work on it's moyo (Dia. 7). AlphaGo keeps pushing Lee Sedol's group and ends in sente, then builds further on it's moyo at the bottom side (triangle in Dia. 8). 


Dia. 8:  Game 5, after black 91 (circle, Lee Sedol is black) 
In a complex middle game Lee Sedol is forced to find efficient ways to prevent AlphaGo from cashing in it's entire moyo. The question is whether AlphaGo can make enough points in the center to outweigh Lee Sedol's already secured territory in the corners (Dia. 8).

Black has about 70 points secured territory, white has about 30 points together in the upper left corner and it's center moyo. Taking 7.5 komi into account this means that white can only win this game if the program gets at least 35 points in the bottom left corner and edge (without giving black any extra compensation for this). Alternatively, white has to find ways to collect additional points in the center. This crude estimate suggests that the game in this position is still open and undecided.


Dia. 9:  Game 5, after white 122 (circle, Lee Sedol is black) 
The game is developing slowly as Lee Sedol first somewhat contains AlphaGo's center moyo, and then creates  a group in the bottom left (Dia. 9). With white's calm responses initially some strength is built though it is unclear what white does want with this. After AlphaGo's extension (circle in Dia. 9) it becomes clear, however, that Lee Sedol has to make a base with his group at the bottom side (or must connect). And that white is trying to lock up black.


Dia. 10:  Game 5, after white 136 (circle, Lee Sedol is black) 
Black connects his group underneath with the right corner and gains some extra points while white builds some force and meanwhile significantly reduces black's influence at the lower right. Then, AlphaGo plays a magnificent split move and nice counter attack (circle in Dia. 10) that adds burden on black and let's white stones working together optimally. 

It is difficult for black to decide what is the best strategy in this position. Lee Sedol can't afford to give up his center stones and at the same time has to minimize white's center moyo as much as possible. AlphaGo's strength at the lower side is being used primarily to keep pressure on Lee Sedol with a range of possible cut actions.


Dia. 11:  Game 5, after black 183 (circle, Lee Sedol is black) 
A comparison between Dia. 10 and Dia. 11 shows that black succeeded to neutralize both white's potential on the left and a significant part of white's center. However, in exchange white gets about 15 secure points as compensation. Whether these are enough for AlphaGo to win the game is unclear, especially as Lee Sedol has still opportunities to further reduce white's potential (both on the left side and in the center). The difference is estimated to be a few points at most for the benefit of AlphaGo (Dia. 11). So, despite Lee Sedol's great moyo reduction skills, it is still deep learning AlphaGo that is ahead due to the sufficient compensation it got along the way. 

What complicates this position is that black's territory is almost defined while white has opportunities to score some additional points at different locations on the board. Lee Sedol's only chance is to restrict AlphaGo's potential as far as possible with smart and effective reduction moves.

For the top Go-profs who comment live on this game, it is hard to say where Lee Sedol perhaps played a lesser move. Even so, AlphaGo managed somehow to get things right after it misjudged Lee Sedol's tesuji earlier in the game. Overall, the flow of the game as well as the way of play of both players looked very well-balanced, constructive and profound.


Dia. 12:  Game 5, end position after white 280 (Lee Sedol is black) 
In the next 100 moves, Lee Sedol unfortunately is not able to catch up his disadvantage of about 2 points. A few minutes before the game is finished, Lee Sedol leaves the Go-board when AlphaGo plays a move that is clearly non-optimal: a sign that the program estimates it will win this game anyhow. Despite the fact that Lee Sedol is ahead on the board, he would loose with about 2.5 points, taking into account AlphaGo's komi (7.5 points).

When he returns, Lee Sedol plays another couple of moves and then resigns after move 280 (final position, Dia. 12) with just less than a handful of endgame moves left. It is the first time in this match that a game has been played until so late in the endgame. And has ended with such a small difference in points.



So this has been another amazing, inspiring, and historic game in which differences in playing strength between the world's top Go-prof Lee Sedol and deep learning program AlphaGo were hard to detect.

With this result, the final outcome of the match is: AlphaGo defeats Lee Sedol by 4 - 1. A result that only a small minority (< 10-15%)  of the more than 280 million people worldwide who watched this match online, would have predicted in advance. AlphaGo has impressed all Go players worldwide with rock-solid, deep reading, sometimes unexpected and really wonderful, effective moves in these games.




During the post-match press conference (never seen so many press and media gathered together), Lee Sedol said: "I am sorry because the match comes to an end". And answering a question about whether the five games might have changed his understanding of the game of Go, Lee Sedol responded: "Basically, I don't necessarily think that AlphaGo is superior to me. I believe there is still more that a human being can do to fight against the AI program. That's why I felt a little bit regrettable because there is more that a human could have shown during this match".




Hassabis stated afterwards: "I am speechless about this most exciting experience we have had so far. AlphaGo made a mistake early in the game because it missed a tesuji. But got itself back into the game. This is by far the most exciting and stressful one of all the games of the match". 

It took Lee Sedol about four games to slightly figure out AlphaGo's way of play: the first two games Lee Sedol probably lost by making a bad strategy decision, the third game Lee Sedol lost due to a fatal mistake already early in the opening, and with the knowledge of the fourth game, it is highly probable that Lee Sedol would have followed other tactics. And would have had a significantly greater chance of winning this match.

A more extensive report of the press conference after the end of the fifth and last game of this match, with more detailed reactions of Lee Sedol and Demis Hassabis, will follow later.

In this fifth and exciting last game of the Google DeepMind challenging match, deep learning AlphaGo played an impressive and very balanced moyo-building game. Even though Lee Sedol had substantial (secure) territory already early in the game and though he was able to thwart most of AlphaGo's moyo plans, the program succeeded in getting enough compensation along the way to stay ahead by a margin of just a few points. And to maintain this small advantage during the second half of the game.


[Part 11: Reactions worldwide and Prize giving]

2 comments:

  1. Review of Game 5 of the Artificial Intelligence Match of the 21st Century: AlphaGo unfamiliar with common tesuji in ultimate moyo game

    This review of the fifth and last game of the Google DeepMind challenging match between deep learning AlphaGo and top Go-prof Lee Sedol (9p) is a highlighting game commentary and analysis including short explanations and discussions of the most important moves and positions, many diagrams, images of the match, and very brief commentaries by top Go-profs and Lee Sedol himself.

    Lee Sedol avoids that AlphaGo will foreclose the bottom right corner and extends his corner (Dia. 5). AlphaGo calculates that it's three stones have enough aji to make a fight meaningful. The program apparently is unfamiliar with a common tesuji (also known as the 'tombstone squeeze'; where you offer two stones, subsequently throw in another one, in order to rob the opponent efficiently from the inner liberties).

    At this point in the game, Demis Hassabis tweeted: "AlphaGo made a bad mistake early in the game (it didn't know a known tesuji) but now it is trying hard to claw it back... nail-biting". He added later: "The tesuji itself appears quite often in pro game records, thus it is most likely learned through the 'policy' network. I personally feel AlphaGo should be aware of the tesuji itself, but Lee Sedol's follow-up move was probably better than expected by AlphaGo". He bites his nails and hopes that AlphaGo very soon gets a chance to make things right again (Dia. 5).

    During the post-match press conference (never seen so many press and media gathered together), Lee Sedol said: "I am sorry because the match comes to an end". And answering a question about whether the five games might have changed his understanding of the game of Go, Lee Sedol responded: "Basically, I don't necessarily think that AlphaGo is superior to me. I believe there is still more that a human being can do to fight against the AI program. That's why I felt a little bit regrettable because there is more that a human could have shown during this match".


    In this fifth and exciting last game of the Google DeepMind challenging match, deep learning AlphaGo played an impressive and very balanced moyo-building game. Even though Lee Sedol had substantial (secure) territory already early in the game and though he was able to thwart most of AlphaGo's moyo plans, the program succeeded in getting enough compensation along the way to stay ahead by a
    margin of just a few points. And to maintain this small advantage during the second half of the game.

    Part 10 of 'The historic match of deep learning AlphaGo vs. Lee Sedol'

    ReplyDelete