Cillian Bracken Conway
28th Jan, 2016

Google has developed a new AI program that is considered a major breakthrough for artificial intelligence. AlphaGo, created by DeepMind, is a board game that uses machine learning to master the ancient and very complex game of Go.

The program achieved what no one has ever done before – beat a top human Go player.

The algorithm has mastered the Chinese board game so intuitively and thoroughly that it has beaten the reigning European Go champion Fan Hui.

DeepMind, the British artificial intelligence company that Google acquired in early 2014, said that the breakthrough is even bigger than the Deep Blue project of IBM, which defeated Russian chess master Garry Kasparov in a 1997 match. It is even bigger in comparison when the Watson computer competed in Jeopardy in 2011 and won.

According to researchers, AlphaGo is more difficult to program for several reasons. In a paper published in Nature, Julian Schrittwieser, one of the authors, outlined the many challenges that the game of Go presented.

  • It involves thousands of possible positions and configurations – 10171 against 1047 in chess.
  • It has a huge number of moves
  • It has an even bigger number of moves per game

In a Google post, researchers explained that building the system involved creating deep neural networks because traditional methods won’t do. They then trained the neural networks on 30 million moves based on the games played by humans who are experts in Go. The training continued until the system was able to predict human moves 57% of the time.

With the system equipped with what are considered good and bad moves, it played against itself to further improve its performance and capabilities.

AlphaGo is a heuristic algorithm, which makes it capable of learning, studying, planning, and making decisions critical to a particular outcome. This is a direct result of researchers not settling for the system to just mimic human players, but also best them.

Further improvements and developments led to AlphaGo discovering new strategies all by itself.

Demis Hassabis, founder and CEO of DeepMind, said that the game “plays different versions of itself…and each time it gets incrementally better”. It is also capable of learning from its mistakes.

He also said that the technology has many potential applications, such as AI-assisted science and health care. The company’s researchers also added that AlphaGo is applicable to a wide range of real world complex data issues and problems. Some of the technology behind the AI system is also believed to appear in some form in the following year or so.

It would take 5 to 10 years before AlphaGo would be deployed broadly and commercially, according to researchers. For now, its next move is to play against Lee Sedol, South Korea’s top Go world champion.

Go is a board game that uses black and white stones set on gridded board. The idea is to surround an opponent’s piece, trapping it to achieve victory. The rules are simpler than chess, but the possibilities are almost endless.