The 'Go'-ahead for Artificial Intelligence
In 1996 chess master Garry Kasparov was beaten by Deep Blue, a supercomputer developed by IBM. Before Kasparov was defeated, many commentators thought that it would be impossible for a computer to beat a human at the game. Chess is a sophisticated game, lauded for its complexity and often used as a measure of human intelligence and human success at the game depends on reading one’s opponent and planning. However, unlike a human, a chess computer is able to analyse all potential moves, a tactic that ultimately led DeepBlue to beat Kasparov.
Fast forward 20 years and there has been another surprise victory for artificial intelligence; this time with a Google-developed program called AlphaGo beating Lee Sedol, a 9thdan champion of the strategy game Go.
Go has long fascinated mathematicians and computer scientists. Back in 1965, the cryptologist I. J. Good described the difficulties involved in a computer beating a human Go player:
“Go on a computer? – In order to programme a computer to play a reasonable game of Go, rather than merely a legal game – it is necessary to formalise the principles of good strategy, or to design a learning programme. The principles are more qualitative and mysterious than in chess, and depend more on judgment. So I think it will be even more difficult to programme a computer to play a reasonable game of Go than of chess.”
Unlike chess, which is played on a board consisting of a twelve by twelve grid with only twenty-four pieces, Go is played on a 19x19 grid and uses counters known as stones. A standard Go set contains a whopping 181 black stones and 180 white stones. In order to win, a player must capture an opponent’s stones. Unlike chess, this is achieved by surrounding a stone with multiple stones which in turn makes mapping the number of potential moves much more difficult if not mathematically impossible. In other words Go is like chess on steroids.
Even as recently as 2015, the best Go programs only managed to reach amateur level and prompted prominent investors such as Elon Musk to comment that we were still 10 years away from a victory against a top ranking professional player.
So how did the computer finally beat a human?
Put simply, by being more human. AlphaGo’s algorithm uses machine learning, neural networks and tree search techniques to make decisions. The program’s neural networks were initially trained to mimic expert human gameplay using data from historical games played by experts. This data consisted of around 30 million moves from around 160,000 games. After this period of learning, the program was trained further by playing large numbers of games against other versions of itself. Once it had reached a certain degree of proficiency, it was trained further by being set to play large numbers of games against other instances of itself. Unlike chess programs, AlphaGo doesn't use a ‘database’ of moves to play.
Because of this, the outcome of the game against Sedol was a complete surprise to AlphaGo’s creators. One of the developers commented:
“Although we have programmed this machine to play, we have no idea what moves it will come up with. Its moves are an emergent phenomenon from the training. We just create the data sets and the training algorithms. But the moves it then comes up with are out of our hands—and much better than we, as Go players, could come up with.”
Predictive coding: science fact not science fiction
The story of AlphaGo gives a fascinating insight how artificial intelligence is developing and uses similar techniques to our own predictive coding technology. In the same way AlphaGo uses machine learning to learn from past games, expert human reviewers train our platform to identify which documents are relevant and make legal document reviews more efficient.
For AlphaGo, this has resulted in a victory against one of the finest human players whereas predictive coding technology has received judicial approval stating that it is as efficient as traditional document review using keyword searches also uses machine learning techniques to mimic a human reviewer.
Many people still have doubts about using predictive coding technology but we hope intelligence victories such as AlphaGo’s will lead to greater public awareness about the capabilities of machine learning technology. After all, if a computer can beat a human in a game as complex as Go, suddenly believing in the capabilities predictive coding seems less of a leap of faith.