Pages

Friday, April 22, 2016

Part 6: Review of Game 1: Lee Sedol underestimates AlphaGo's incredible fighting power (The historic match of deep learning AlphaGo vs. Lee Sedol)


[NL Versie]

Review of Game 1: Lee Sedol underestimates AlphaGo's incredible fighting power


The first game of the Google DeepMind challenging match  between deep learning AlphaGo and top Go professional Lee Sedol (9p) is the most important game of this match as this is the first time an AI program plays without handicap against one of the strongest human Go-players of the world.


AlphaGo has not learned from professional games and therefore, as repeatedly stated and emphasized by Demis Hassabis, has not a single game of Lee Sedol in it's database. Lee Sedol doesn't know much more about the program than what he has seen from AlphaGo's match against Fan Hui (2p) last October: a strong but now and then malfunctioning program, sometimes making obvious mistakes in complex situations, without too much whole-board understanding. And that against a substantially weaker Go-player who made many overplays Lee Sedol probably never would make. 

Both sides will thus be scanning each other for the very first time, finding out the strategies, way of thinking, patterns, playing strength of the opponent and exploit any possible weaknesses. While AlphaGo originally trained on 130,000 amateur games (up to 8-9 dan = ~1-2p) and further improved from millions of self-play games, Lee Sedol played several thousands of top-tournament games in the almost 25 years of his professional Go-life. 





The opening moves of the first game in this historic match between mankind and AI are played while it is early in the morning 5 AM in the Netherlands. Here the match is watched under extremely high-voltage excitement in the European Go Cultural Center (EGCC) by more than 60 people that have come from all corners of The Netherlands with or without sleeping bag, to experience this once-in-a-lifetime event together real-time and to live through the game as conscious as possible. 

Two big screens are placed on the wall from which the online game commentary and playing room of the match are displayed. Huge posters, a demonstration board and many laptops with all-sides discussions of the match worldwide. Meanwhile the game is discussed by the strongest European players we have: Merlijn Kuin (6d), Peter Brouwer (6d), and Guo Juan (5p, inactive as pro player).

It is no secret that Lee Sedol consulted several Go-players and programmers in preparation to this match. This may be also one of the reasons why Lee Sedol will test in specific manners how AlphaGo will treat and react on moves that the program never has seen (or learned) before. Besides, Lee Sedol is assisted by a team of psychologists and coaches to deal with the enormous mental pressure and media attention worldwide. And to let himself not being brought out-of-balance by the non-ability to assess the emotional state and mental stability of his opponent (in this case, the biggest challenge is actually to win from himself in particular). 


After Lee Sedol's move with black 7 (see Dia. 1) Merlijn Kuin (6d) kicks off immediately: “this is a rather unique move and very aggressive of black. According to me, this is a position which appeared never before on the board. I am convinced that Lee Sedol uses a  preconceived strategy that fits his style". Lee Sedol makes a bad decision by choosing an unusual opening of which he is sure it is absent in AlphaGo's database. However, Lee Sedol himself is also unfamiliar with such a rare and less optimal opening pattern. 


Dia. 1:  Game 1, after white 14 (lowest triangle, Lee Sedol is black).
Black's move 7 is marked with a green dot. 
When seeing the first moves played by AlphaGo, Guo Juan (5p) shakes her head in a clearly disapproving manner (Dia. 1): "this program really needs a good teacher who will knock and bestow upon the head when playing such a bad move, every just knows that this move isn't good at all ... these first white moves are obviously not optimal and look like beginner's mistakes". The order in which the white moves are played (marked with a triangle in Dia. 1), provide black with the opportunity to answer as in the game. Therefore, the result is somewhat less for AlphaGo (see Dia. 1).


Dia. 2: 
 Game 1, after black 23 (circle, Lee Sedol is black) 
With Lee Sedol's move 23 (circle in Dia. 2) the game is going up in flames immediately. Guo Juan (5p) comments almost instantly: "this move by Lee Sedol is much too compulsory and forcing" and she indicates that black's tsuke (touching move) is a huge overplay. The tone of the game has been set as the rest of the game flow will be determined by Lee Sedol's huge underestimation of the program. AlphaGo cuts at once and takes the initiative by putting Lee Sedol under high pressure.

After barely 25 moves (Dia. 3), there are seven groups fighting each other. Peter Brouwer (6d) remarks that the shapes and patterns in this game are not particularly beautiful. And Michael Redmond (9p) comments: "in reaction to a slightly less optimal move by AlphaGo, it appears that Lee Sedol comes at once with an overplay. This human reaction is well understandable and makes AlphaGo at least one stone stronger". 


Dia. 3:  Game 1, after white 50 (circle, Lee Sedol is black) 
With move 48 (triangle in Dia. 3), white attacks the black group in the upper right-hand side. Black answers (square in Dia. 3) and adds some strength to his group by creating additional eye potential. In case of emergency, black can connect underneath with the three black stones at the top right side (e.g. by black R18). However, since it is far too early for black to play himself on 48 (triangle in Dia. 3), white's move 48 is seen by top profs as a mistake by AlphaGo (even though it is sente).


Dia. 4:  Game 1, after white 80 (white stone with square)  
After AlphaGo pushes from behind on the seventh line in Dia. 4, it becomes clear that the program does not respect widely accepted Go-proverbs ('in the far east a child gets a smacking or a hit with a japanese fan on his hands"). 

With AlphaGo's strategy Lee Sedol gets huge influence in the center. But at the expense of a much weaker corner at the bottom right. After black neutralizes most of the aji of the two white cutting stones in the center (triangle in Dia. 4), white strikes and invades  (circle in Dia. 4). Evidently, AlphaGo wants to use the force of the white wall and to transform that into fighting power.

Lee Sedol plays tenuki and attacks the white stone in the bottom left corner while expanding his center moyo (blakc stone with square in Dia. 4). In turn, AlphaGo plays tenuki as well and captures two black stones, thereby removing most dangers (aji) around the center position (white stone with square in Dia. 4).

However, since removing black's aji is not urgent right now and with this move AlphaGo makes an important error in positional judgement. After the program's blunder, top prof Gu Li (9p) stated: "Lee has now a chance of 90% to win this game". And Michael Redmond comments: "AlphaGo is very accurate and profound in computing moves and very balanced in influence and territory. After an earlier better position for AlphaGo, Lee Sedol has brought the game in balance again".


Dia. 5: 
 Game 1, after white 102 (circle, Lee Sedol is black)
And then suddenly in a calmly developing game, a dramatic upheaval happens: "AlphaGo's move 102 (Dia. 5) really is a superhuman moveaccording to Michael Redmond. It looks like white has carefully prepared this invasion with moves in the upper right and with the corner approach --25 moves earlier-- at the bottom right (white 78, circle in Dia. 4). 

A top Go-prof who thoughtfully suggested this move was laughed at in his face: initially many Go-profs are rather negative and demand proof for this move. AlphaGo just plays it as nothing special. In second thought, however, profs worldwide admire this wonderful effective invasion. Others are heavily  disconcerted by this remarkable move of the AI program as it shows it's incredible strength and profound 'understanding' of the concepts the game (including aji). This clearly shows how different this version of AlphaGo is compared to that during the Fan Hui match. 

With this invasion AlphaGo attempts to blow up entirely Lee Sedol's position at the right side of the board in a terrifying fight (circle Dia. 5) while using the great influence it has built earlier. At the same time, this invasion gives opportunities to make a base for white's group by exploiting the weaknesses in black's position on the right-hand side (note how the white group extends over the full length of the board, see Dia. 6).


Dia. 6:  Game 1, after white 116 (circle, Lee Sedol is black)
With AlphaGo's impressively strong and fabulous move, Lee Sedol has no choice other than to give up is three stones in the upper right (see Dia. 6). In exchange, he gets three white stones but black's potential at the right side has disappeared like frost under the morning sun (in gote). Then, AlphaGo plays a rather slow (and a bit cowardly as it seems unnecessary right now and is not optimal according to the commentators) but surely effective move which secures the white corner in the upper left (circle in Dia. 6).

To play this move right now in the game (while there are several other, larger moves) indicates that the program believes it is ahead and has taken a sufficient lead to win the game. AlphaGo is absolutely not interested in playing the most optimal, largest or most efficient moves: the program only plays moves that provide the highest probability of winning the game.


Dia. 7:  Game 1, after black 127 (circle, Lee Sedol is black)
Anyhow, Lee Sedol starts a fight in the bottom right corner in an attempt to put AlphaGo under pressure (Dia. 7). Due to some mistakes by black,  white succeeds in making a living group that is much larger than what black should have allowed. Lee Sedol seems already too much behind to catch up in the rest of the game.


Dia. 8:  Game 1, after white 136 (circle, Lee Sedol is black) 
Nonetheless, Lee Sedol checks AlphaGo's whole-board awareness (triangle in Dia. 8): does white notice that his long, elongated group is not yet alive? With the effective defense by AlphaGo it becomes clear that the program has excellent whole-board understanding (circle in Dia. 8).


Dia. 9: Game 1, final position after white 186 (circle, Lee Sedol is black)
After about 50 endgame moves, Lee Sedol resigns. With AlphaGo's move 186 (circle in Dia. 9), Lee Sedol is more than 5 points behind. No doubt, Lee Sedol is greatly saddened by his severe underestimation of the mind-blowing strength and amazing fighting of this version of AlphaGo (as opposed to that during the Fan Hui match about half year ago).





This is the first time in history that an AI program defeats a top Go-prof in a formal game without handicap. Go-profs worldwide are speechless and all agree that this has been a fabulous and historical game. And of course, there is a lot more to say in detail about the firm attacks, remarkable, aggressive and deliberate moves, tactics and strategies of both sides in this game. 

The era in which a human could wipe an AI program (without handicap) from the board is past and once and for all closed. This magnificent and superbly profound way of playing Go is never shown before on our planet. And will leave behind incredibly deep imprints and heavy tracks on the hundreds of millions of people worldwide who were following this game online.


Lee Sedol loses this first game against AlphaGo and reacts deeply saddened, disappointed and very touched: "I was very surprised because I did not think that I would lose the game. I am shocked by how well AlphaGo played, I can admit that. I didn't think that AlphaGo would play the game in such a perfect manner. A notable mistake from my side at the beginning of the game kept continuing during the game and lasted until the very last. AlphaGo's early strategy was excellent and I was stunned by one unconventional move Alphago made (see Dia. 5that a human never would have played. I am in shock, but what's done is done".

Despite his loss in this opening game, Lee Sedol stated he did not regret accepting the challenge: "I had a lot of fun playing Go this game and I'm looking forward to the future games". 


Michael Redmond (9p): "In this first game of the series, AlphaGo triumphed by a very narrow margin. While Lee Sedol had led for most of the match, AlphaGo managed to build up a strong lead in the closing stages of the game. AlphaGo played profound and solid moves with which it secured victory and ultimately has punished Lee Sedol severely and consequent for his lesser opening move (move 7) and aggressive overplay (move 23)". 


In this first game, AlphaGo has shown evidently  that it has a taste for aggressive and offensive play, something the program not necessarily demonstrated during the Fan Hui match. In this game AlphaGo has exhibited surprisingly accurate, profound, and versatile judging of the position.

And AlphaGo has played at least one incredibly wonderful move that achieved so many goals at the same time that it is hard to apprehend (Dia. 5). During the fight that followed, the program succeeded in getting enough advantage so that it was confident to secure the upper left corner with an all-telling move and claiming the victory of the game (Dia. 6).



Although Lee Sedol controlled the majority of the game, AlphaGo managed to rebuild itself in a phenomenal way and secured a clear advantage during the last 20 minutes of the game after which Lee Sedol resigned.

With all his hubris and self-assurance prior to this match, his underestimation of the startling, mind-boggling, fabulous, and wonderful play and strength of AlphaGo, his expectations on the basis of the games during the Fan Hui match, his immense concentration and involvement with this first game of the match, and the enormous media attention and psychological pressure by all eyes of the world, this extremely surprising and unexpected loss must have felt particular anguish for Lee Sedol.

2 comments:

  1. This comment has been removed by the author.

    ReplyDelete
  2. Review of Game 1: Lee Sedol underestimates AlphaGo's incredible fighting power

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

    This first game is also the most important game of the Google challenging match: for the first time in history an AI program dares to play in formal games without handicap against the strongest Go-player of the world over the last fifteen years.

    With all his hubris and self-assurance prior to this match, his underestimation of the startling, mind-boggling, fabulous, and wonderful play and strength of AlphaGo, his expectations on the basis of the games during the Fan Hui match, his immense concentration and involvement with this first game, and the enormous media attention and psychological pressure by all eyes of the world, this extremely surprising and unexpected loss must have felt particular anguish for Lee Sedol.

    Part 6 of 'The historic match of deep learning AlphaGo vs. Lee Sedol'

    ReplyDelete