20 years after DeepBlue became the first artificial intelligence to defeat a chess world champion by winning a game against Kasparov, it’s successor, DeepMind, made another giant leap in the history of AI by defeating the world’s top Go player Lee Sedol.
An historic victory
The 9th March 2016, the Google-backed AI project DeepMind won it’s first game against the 9-dan, world top professional Go player Lee Sedol.
The event is a 5 games match hosted in Seoul, South Korea, and broadcasted live on YouTube. The winner of this best-of-five will receive a prize 1 million dollars.
DeepMind’s AlphaGo program already won a match 5-0 against the European champion Fan Hui in October 2015. But Lee Sedol is on another level, ranking at 9 dan and being the second best Go player in the world, while Fan Hui is only a 2 dan player.
AlphaGo algorithm is run on Google US-based servers, displaying the computer’s moves on a screen. Alpha Go‘s moves are then reproduced on the board by a human operator.
The first game happened the 9th March and saw the victory of DeepMind after a 3 and a half hour game. Lee Sedol stated after the game that he was surprised by how strong the AI was, but that he made a major mistake in the early game and was looking for the next game as he thought he still had his chances to win.
However, the 10th March, the second match ended in a second victory for DeepMind, with Lee Sedol declaring that he was “speechless” and admits “A very clear loss[…]. From the very beginning of the game I did not feel like there was a point that I was leading”.
The game of Go
This victory is so important because for a long time, the game of Go have been an insolvable problem for AIs. This is because, unlike chess, the game of Go is based on the intuition and the experience of the player.
The basic principles of the Go are rather simple : 2 players face on a 19*19 lines grid and places stones of color (black and white) at the intersections of this grid.
The goal is to control a greatest area of the board than your opponent (having more stones on the board than him to simplify). Opponent’s stones can be removed from the board by encircling them.
This game is said to be much more deep and complex than chess, and even if classic moves and tactics exist, it’s not as rigid as chess since the number of possible configurations is too great to memorize all the possible moves.
How is AlphaGo built ?
Computers have been in trouble playing Go because the size of the board and the liberty allowed by the rules create a number of possible configuration way beyond our bruteforce calculation power at this day.
This is why AlphaGo uses a combination of machine learning, complex tree-search algorithms implemented with neural network inspired structure.
The program is able to learn and become a better Go player over time, like and human player. It have used the data from millions of moves from professional games to reinforce it’s algorithm, and then trained further by playing multiple games against himself.
We don’t really know what is the hardware processing the program, but since it’s running from Google servers, we can say that it packs a lot more computational power that the average PC.
Apart from playing Go, the DeepMind project also aim to use AI to solve more serious problem as helping identifying patients at risk, and globally to take the Artificial Intelligence further and further. And they already did !
UPDATE (13 March 2016) : Although AlphaGo won the 3rd game, Lee Sedol managed to outplay the AI in the 4th game and won the first game of this match !
Sources :
G2 blog (blog) 
Finally ! The human is not completely out-matched ! I thought Lee would win the first and maybe the second match, then the AI will adjust its algorithms and it would become impossible for the human player to win. But the fourth match proves that both are a similar level.
It’s also interesting in term of strategy, because the computer discovers new ways to play. This experience will definitely enrich the world of Go.
LikeLike
I find it really impressive all the progress that we have made in terms of AI. I think this domain will evolve even more in the next few years ! That’s really neat !
LikeLike