In November of 2019, Lee Sedol, once the world’s champion of Go, announced his early retirement.
Even if I become the number one, there is an entity that cannot be defeated.
Lee Sedol
In 1997, a computer defeated the reigning chess champion by sheer evaluation power. Churning through millions of possible plays and counterplays in seconds, it was able to account for a broad swath of possible futures in the time it took to pick a single move.
In response, chess fanatics created a new game called “Arimaa,” that would present so many possibilities even a supercomputer couldn’t churn through them all. They forgot that such a game had already existed for four thousand years.
A 19×19 Go board with three possible states for each of its 361 spaces (empty, white, black) has 3361 possible configurations. Not all of these are legal in a real go game, but the approximate number of legal possibilities has been known since 2006: 2.08 * 10170 positions. The exact number was discovered by Jon Tromp in 2016. Instead of saying words like “ten million vintigillion googol,” let’s provide the number of atoms in the universe as reference. Estimated at between 1078 and 1082, the known universe contains a miniscule fraction of atoms compared to the set of possible go games. That protected human champions like Sedol for about nineteen years.
What took him down was another technique from the 1950s, again applied to modern technology. Based on the “Law of Effect” observed in animal behavior, Trial and Error learning sought to duplicate the process by which animals learn to seek situations that lead to rewards and avoid those that lead to penalties. By the 1980s, this field had combined with ideas from dynamic programming to become modern reinforcement learning.
By the late 1990s, Yann LeCun’s LeNet demonstrated the usefulness of a convolutional neural network for recognizing handwritten digits. A convolutional neural network convolves an image, applying a single module to a fixed-size window that operates on each valid position within an image.
In the handwritten digit task, an image is represented to a computer as a fixed-size grid of pixels, each with a single value representing its brightness.
Does this sound applicable to another task?
Furthermore, hardware continued to advance. The success of the video game industry had pushed the state of the art in graphics technology for decades. Graphics processing units (GPUs) had grown to apply complex transformations to images in real time using operations similar to those in use in nascent neural networks. In the early 2000s, researchers learned to harness one for the other, increasing neural network processing speeds by up to fifty times or more. Although companies such as Google have invented proprietary alternatives specifically for neural network processing, GPUs remain the primary device for processing neural networks in research and industry.
In 2016 a reinforcement learning convolutional neural network system using 280 GPUs took on Lee Sedol and won four games out of five. Like Deep Blue, it simulated possible future moves, but with its neural network reinforcement learning algorithms, it narrowed the impossible search space to something more manageable.
Unlike the mechanical Turk, this time there was no doubt that AlphaGo was completely mechanistic. Nevertheless, it had conquered uncertainty in a realm where brute force alone could not. In 1836, Edgar Allan Poe insisted that to conquer uncertainty such an automaton must be “regulated by mind.” In the absence of the human mind hidden inside the Turk’s box, could it be that AlphaGo was indeed an artificial equivalent?
This question may be a debate for philosophers, but one pair of moves between Sedol and AlphaGo in the second game of their fateful match, and the way AlphaGo and Sedol responded to it, may change how people approach the debate in the future. Tune in next week.
I take requests. If you have a fictional AI and wonder how it could work, or any other topic you’d like to see me cover, mention it in the comments or on my Facebook page.
Samuel Munk 