A striking example of machine creativity is DeepMind’s AlphaGo, which first beat European Go champion Fan Hui in October 2015 and then Go world champion Lee Sedol 4–1 in a match in March 2016. The ancient board game of Go has long been a milestone in AI research due to it being one of the most complex games humans have designed. Chessplaying AIs can simply use brute force and search all potential moves until the winning ones are found, but this is not possible in Go since there are more possible configurations of the board, and thus potential moves, than there are atoms in the universe. Until AlphaGo proved otherwise, the game was seen as requiring human-level pattern recognition and something akin to intuition.<sup>11</sup> To learn how to play Go, AlphaGo relied on a combination of deep neural networks and Monte Carlo tree search. Its architecture consists of three convolutional neural networks (CNN) that were divided into one ‘value network’ and two ‘policy networks’.<sup>12</sup> The CNNs were trained on image data that represented individual moves (‘states’) from 160,000 games recorded from top Go players (Silver et al. 2016). The value network learned how to play the game and to estimate the probability of winning given any current state of the game, while the policy networks learned what move to select given the current state of the game. In turn, the two policy networks’ decision-making was reinforced by having them play each other 30 million times and keeping track of the best outcome of the games (Silver et al. 2016).
AlphaGo’s creativity or intuition was evident in the 37th move of the second game against Sedol. Although this move was later described as ‘creative’, ‘unique’ and ‘beautiful’, as well as ‘inhuman’ by professional and champion Go players (including Sedol), it at first took everyone by surprise (Metz 2016). After AlphaGo made this move, Sedol took an unusually long time to respond. Commentators noted how strange it was, some thinking that the neural network had made a mistake because a human would never make such a move. But the move was deliberate and based on what AlphaGo’s policy networks had learned. The engineers at DeepMind could not have predicted this move and could only determine AlphaGo’s reasoning after it had made its move: having calculated the probability of a human making this particular move at that particular state of the game at 0.010 per cent, it nevertheless made it (Metz 2016). AlphaGo arguably made a plan and executed it. Not only did AlphaGo learn how humans play, but it went beyond the constraints that generations of human Go players had placed on how to play the game, making seemingly counterproductive and highly novel moves, and playing the game in styles that no human would (Metz 2016). AlphaGo’s creativity comes from moves like this and from how it learned to play Go in a completely new and inhuman style; if a human had done any of this, she would likely have been recognized as original, creative and highly intelligent.
120
inhuman power