His opponent — the computer program AlphaGo, from Google-owned DeepMind — had, in the 37th move of the game, placed its stone in what the Go community calls a “shoulder hit”; a move professional Go players seldom use. Stunned, Lee just walked out of the room. AlphaGo appeared to demonstrate creative initiative exceeding the best human players. Or had it?
Lee returned a few moments later and played a brilliant game, though he still conceded defeat after 211 moves.
Go is a simple game that is hard for computers: It has far too many possible moves for a computer to explore. While no reasonable Go program relies only on trying different moves, until AlphaGo, the auxiliary mechanisms lagged behind human play. An earlier version of AlphaGo had lost just six months prior to Fan Hui, a much less skilled player than Lee. AlphaGo’s eventual victory over Lee (Lee lost the match to AlphaGo, four games to one) was unexpected. Most people expected Lee to win.
How did AlphaGo win? And was its play, as Marcus du Sautoy suggests in a recent article at New Scientist—“True AI creativity is coming and will reveal the minds of machines” — “truly creative”?
Computer expert Dan Maas provides a lay summary of DeepMind’s Nature paper describing AlphaGo; I encourage you to read one or both articles. Because Go’s combinatorial complexity precludes trying out all possible moves, AlphaGo combined a Monte Carlo (that is, random) search tree with two “neural networks”: one to evaluate moves and another to measure the likelihood of a win, given a board arrangement of stones.
While these techniques are new, what made AlphaGo especially effective is the training DeepMind experts gave it. DeepMind first trained the system using past games played by Go masters — giving AlphaGo’s play style a human-like skew. But then DeepMind further trained AlphaGo by having it play over a million games against previous iterations of itself (that is, AlphaGo played AlphaGo). This additional training took it to places, through random selection, that humans seldom ventured into. This new terrain led to its unexpected 37th move — and to a disastrous move that cost it the fourth game against Lee.
In short, when AlphaGo played the “shoulder hit,” it exhibited no more creative insight than when it played pedestrian moves. That is, AlphaGo always selected moves only because prior training data suggested them to be good or useful at that point in the game. The creative initiative resides entirely in the DeepMind team and their insight to let AlphaGo play against itself as a means by which to “learn.”
Our surprise at AlphaGo’s move says more about our inability to predict what a program will do than about any creative effort of the program. We’ve known for decades that we cannot predict the results of any moderately complex computer program. Thus, it’s surprising we were surprised by AlphaGo’s move because there was no real surprise: Complex computer programs will, occasionally, do things we do not expect. Sometimes they work.
So, if AlphaGo’s unexpected move failed to display creativity, what, then, is creativity? Could a sufficiently advanced AI produce creative results? I will explore both of these questions in future posts.
See also: Part II: Why AI fails to actually create things Only one of the traits du Sautoy suggests is an essential part of creativity
Also: Creativity does not follow computational rules