Researchers at Google DeepMind have developed an artificial intelligence program  — AlphaGo — that can outgun a professional player at the ancient Chinese game of strategy, Go.

While it may not sound world changing, artificial intelligence (AI) developers use games to develop and test their algorithms, with the ultimate goal to apply these techniques to important real-world problems such as climate modelling to complex disease analyses, says CEO of DeepMind, Demis Hassabis,

The team published the algorithms used by AlphaGo in the journal Nature on January 27.

Nature Senior Editor, Dr Tanguy Chouard says that this achievement “will surely be seen as a historical milestone in artificial intelligence.”

DeepMind is a London-based artificial intelligence (AI) company co-founded by Hassabis, Shane Legg and Mustafa Suleyman in 2010, and acquired by Google in 2014.

“Go is probably the most complex game devised by man,” says Hassabis, with each move having 10 times more possible outcomes than in chess.

This complexity has proven to be a big obstacle for programmers of AI. Creating a program that can defeat a human expert player has long been considered the Holy Grail of AI developers.

Deep learning

Until AlphaGo beat a three-time reigning European Go champion, Go-playing programs have only been able to attain amateur status.

The DeepMind team previously used deep learning— a form of machine learning where programs model how humans learn— to train a computer to play video games by trial and error without providing any prior instructions; the computer performing better with each successive game to eventually surpass human players.

They published the research in Nature in February 2015.

To create AlphaGo, the research team integrated more than one deep learning technique. Silver says that the key was to reduce the search possibilities to something more manageable by limiting the search to moves most likely to win — teaching the program to think intuitively, more like a human player.

They “taught” the computer thousands of moves used by human Go players and allowed the machine to learn from trial and error by itself. Essentially, they succeeded where others had failed by adding human intelligence to their algorithms.

One of the pioneers of artificial intelligence, Marvin Minsky, passed away only days before the announcement of the success by AlphaGo. Minsky — who helped lay the foundations for artificial intelligence research — also believed that the solutions to the world’s most challenging problems could one day be solved by intelligent machines.
—Sue Min Liu