The race to build a computer program that can beat a professional Go player has been won by the team at Google DeepMind, led by Demis Hassabis. The program, called AlphaGo, won 5-0 in games played in October 2015, and was only recently made public in the journal Nature on January 27, 2016. The human player, three-times European Champion, Fan Hui, is rated as a 5 dan Professional. AlphaGo will now play the world number-one Go player Lee Sedol on March 9 in Seoul, South Korea. Google beat tech rivals, Facebook and Microsoft, to this milestone. These companies, as well as universities worldwide, also have research programs in computer Go. The technology in AlphaGo can also be applied in other areas of machine learning. Computer Go therefore provides a useful microcosm for improving deep learning technology.
Why the computer Go milestone matters
The technology behind AlphaGo is a combination of the well-researched technique Monte Carlo tree search with the newer deep neural networks, a branch of machine learning that has made rapid strides in the past five years, and a specialty at DeepMind, the London-based AI company acquired by Google in 2014. Computer Go is a challenge in AI because it requires machine skills that can learn good play as opposed to the brute-force searching of optimal moves as used in computer chess. Go has a far greater number of possible moves available than chess, and brute-force methods using the best supercomputers cannot produce a good move in reasonable time. The achievement of the milestone is that the technology used in computer Go can be applied in other applications of AI where machine learning and reasoning is required.
Computer chess is also improving
The use of deep learning is also being applied in computer chess. Matthew Lai, while at Imperial College London, created Giraffe, an open source chess engine that uses deep learning techniques. The engine teaches itself through playing many games, and has achieved International Master level. Matthew Lai has, however, discontinued the Giraffe project and joined DeepMind in January 2016.
Michael Azoff, Principal Analyst, Ovum Infrastructure Solutions Group