Artificial Intelligence-2 Human GO Champ-0
We are currently witnessing one of the big milestones in Artificial General Intelligence. An AI has beaten the best human GO player in the world.
As of this writing, 10-year world champion, Lee Sedol has lost the first two games of a best-of-five match against Google’s Deep Mind AI program, called AlphaGo. The match is taking place in Seoul, South Korea from March 9 -15, 2016.
Created in China 2,500 years ago, GO (or Baduk) is one of the oldest board games played today. Its name literally means the Encircling Game and that’s precisely what you have to do. You have to surround the largest total area of the 19×19 board with your playing pieces or stones.
The simplicity of its playing pieces and rules. however, masks an incredible hidden complexity that far surpasses other games like chess or thello. In fact, it is Go’s complexity that has long made it one of the outstanding Grand Challenges for Artificial Intelligence.
So how complex is it actually? Well, let’s look at chess and atoms first.
It is often cited that there are more distinct games of chess than there are atoms in the observable universe. The number for the latter is 10^80 atoms (one hundred quinvigintillion…give or take 3).
For chess, one of the most frequent maximum number of games I come across is 10^120, which is certainly far larger than the atom number. I have to say though that this number is unrealistic as it includes a plethora of legal individual moves that are completely unreasonable. A better approximation is 10^40 (10 duodecillion). This probably contains most of the games reasonable people would play yet it is still a massively huge number.
That’s nothing. GO puts all those numbers to shame. Most who have bothered to calculate agree that GO has a mind-numbing 10^761 possible game scenarios. Read that again…10^761. Did you know there’s a name for that? Of course there is, it’s One hundred duocendoquinquagintillion
It is these vast numbers of possible games that prevent modern computers from using a brute-force method to look many moves ahead of what a human can do when playing GO. This is what IBM’s supercomputer Deep Blue used to look ahead many moves beyond what chess champion Garry Kasparov could do in 1997, ultimately defeating him.
AlphaGo, created by Google’s Deep Mind acquisition, had to do things differently. They had to try and instill something more akin to human “intuition” and “feel” rather than an exhaustive search of possible moves. It started with a technique called Deep learning on its neural network. In this case, that meant exposing the software to massive amounts of data, 30 million moves, from expert players to teach it how to play. That was the first step and by its nature could only bring you so far. Best case, the system would only be as good as those expert players.
To beat the best humans, the next step included what’s called reinforcement learning. This involved the system playing against different iterations of itself, learning which new moves worked and which ones didn’t. This allowed it to learn as it played by analyzing more and more data. The end result was a new suite of plays and strategies that it learned all by itself.
This reminds me of a comment by Gizmodo commenter Vasshu. He, she or it said: “And with all the supercomputing power, it still cannot LEARN how to beat a professional go player.”
In fact, as I just described, learning is exactly how AlphaGo got as good as it is.
Regarding the match, Lee Sedol was initially very confident. He predicted a 5-0 match or 4-1 at worst. After talking with Deep Mind’s CEO and learning about AlphaGo, his confidence had certainly waned. He said,
“I don’t think I can win 5:0…I feel I should be on the edge (during the match)”
After his first loss, Lee said that:
“AlphaGo made some moves that no human would ever make. It really surprised me,”
After his second loss on March 9th, he said the following:
“Yesterday, I was surprised, but today I am quite speechless…Yesterday I felt like AlphaGo played certain problematic positions, but today I felt that AlphaGo played a near perfect game. There was not a moment I felt like its moves were unreasonable.”
Lee certainly has an uphill battle ahead of him. He has to win the next three games to win the match and take home the million dollars. Yet, even if he does, I don’t think it will matter that much. The point has been made. AI has, to a significant extent, conquered the game of GO. If history is any guide, it will soon reign supreme far far beyond what any mere (unaugmented) human can muster. This could lead to far more than just a one-off AI that’s unbeatable at GO. The application of this brand of AI to other fields could be transformative.
For example, Nick Bostrom of Oxford University’s Future of Humanity Institute said:
“AlphaGo is really more interesting than either Deep Blue or Watson because the algorithms it uses are potentially more general-purpose,”
DeepMind founder Demis Hassabis said “It’s a natural fit for robotics,” in that it could help robots interact and respond to their environment. These AI systems could work with scientists to help zero in on research areas most likely to produce the big results everyone hopes for. He continues…“The system could even suggest a way forward that might point the human expert to a breakthrough.”
Other areas beyond scientific research and robotics could also be impacted, like Digital assistants and financial investments, to name just a couple.
Chris Nicholson, founder of the deep learning startup Skymind, really put it into perspective when he said:
“You can apply it to any adversarial problem—anything that you can conceive of as a game, where strategy matters,”
There’s still far to go for a truly artificial general intelligence. This latest attempt at machine learning certainly seems to have promise.
Still…..It can’t hurt to start practicing your robot sucking up techniques.
Image Credit: https://gogameguru.com/i/2016/01/DeepMind-AlphaGo.jpg