so that as the Go-playing program played more and more games it would learn from its mistakes.
Hassabis had learned about a similar idea implemented by the artificial-intelligence researcher Donald Michie in the 1960s. Michie had written an algorithm called ‘MENACE’ that learned from scratch the best strategy to play noughts and crosses. (MENACE stood for Machine Educable Noughts And Crosses Engine.) To demonstrate the algorithm, Michie had rigged up 304 matchboxes representing all the possible layouts of noughts and crosses encountered while playing. Each matchbox was filled with different-coloured balls to represent possible moves. Balls were removed or added to the boxes to punish losses or reward wins. As the algorithm played more and more games, the reassignment of the balls eventually led to an almost perfect strategy for playing. It was this idea of learning from your mistakes that Hassabis wanted to use to train an algorithm to play Go.
Hassabis had a good model to base his strategy on. A newborn baby does not have a brain that is pre-programmed to cope with making its way through life. It is programmed instead to learn as it interacts with its environment.
If Hassabis was going to tap into the way the brain learned to solve problems, then knowing how the brain works was clearly going to help in his dream of creating a program to play Go. So he decided to do a PhD in neuroscience at University College London. It was during coffee breaks from lab work that Hassabis started discussing with a neuroscientist, Shane Legg, his plans to create a company to try out his ideas. It shows the low status of AI even a decade ago that they never admitted to their professors their dream to dedicate their lives to AI. But they felt they were on to something big, so in September 2010 the two scientists decided to create a company with Mustafa Suleyman, a friend of Hassabis from childhood. DeepMind was incorporated.
The company needed money but initially Hassabis just couldn’t raise any capital. Pitching on a platform that they were going to play games and solve intelligence did not sound serious to most investors. A few, however, did see the vision. Among those who put money in right at the outset were Elon Musk and Peter Thiel. Thiel had never invested outside Silicon Valley and tried to persuade Hassabis to relocate to the West Coast. A born-and-bred Londoner, Hassabis held his ground, insisting that there was more untapped talent in London that could be exploited. Hassabis remembers a crazy conversation he had with Thiel’s lawyer. ‘Does London have law on IP?’ she asked innocently. ‘I think they thought we were coming from Timbuctoo!’ The founders had to give up a huge amount of stock to the investors, but they had their money to start trying to crack AI.
The challenge of creating a machine that could learn to play Go still felt like a distant dream. They set their sights at first on a seemingly less cerebral goal: playing 1980s Atari games. Atari is probably responsible for a lot of students flunking courses in the late 1970s and early 1980s. I certainly remember wasting a huge amount of time playing the likes of Pong, Space Invaders and Asteroids on a friend’s Atari 2600 console. The console was one of the first whose hardware could play multiple games that were loaded via a cartridge. It allowed a whole range of different games to be developed over time. Previous consoles could only play games that had been physically programmed into the units.
One of my favourite Atari games was called Breakout. A wall of coloured bricks was at the top of the screen and you controlled a paddle at the bottom that could be moved left or right using a joystick. A ball would bounce off the paddle and head towards the bricks. Each time it hit a brick, the brick would disappear. The aim was to clear the bricks. The yellow bricks at the bottom of the wall scored one point. The red bricks on top got you seven points. As you cleared blocks, the paddle would shrink and the ball would speed up to make the game play harder.
We were particularly pleased one afternoon when we found a clever way to hack the game. If you dug a tunnel up through the bricks on the edge of the screen, once the ball made it through to the top it bounced back and forward off the top of the screen and the upper high-scoring bricks, gradually clearing the wall. You could sit back and watch until the ball eventually came back down through the wall. You just had to be ready with the paddle to bat the ball back up again. It was a very satisfying strategy!
Hassabis and the team he was assembling also spent a lot of time playing computer games in their youth. Their parents may be happy to know that the time and effort they put into those games did not go to waste. It turned out that Breakout was a perfect test case to see if the team at DeepMind could program a computer to learn how to play games. It would have been a relatively straightforward job to write a program for each individual game. Hassabis and his team were going to set themselves a much greater challenge.
They wanted to write a program that would receive as an input the state of the pixels on the screen and the current score and set it to play with the goal of maximising the score. The program was not told the rules of the game: it had to experiment randomly with different ways of moving the paddle in Breakout or firing the laser cannon at the descending aliens of Space Invaders. Each time it made a move it could assess whether the move had helped increase the score or had had no effect.
The code implements an idea dating from the 1990s called reinforcement learning, which aims to update the probability of actions based on the effect on a reward function or score. For example, in Breakout the only decision is whether to move the paddle at the bottom left or right. Initially the choice will be 50:50. But if moving the paddle randomly results in it hitting the ball, then a short time later the score goes up. The code then recalibrates the probability of whether to go left or right based on this new information. It will increase the chance of heading in the direction towards which the ball is heading. The new feature was to combine this learning with neural networks that would assess the state of the pixels to decide what features were correlating to the increase in score.
At the outset, because the computer was just trying random moves, it was terrible, hardly scoring anything. But each time it made a random move that bumped up the score, it would remember the move and reinforce the use of such a move in future. Gradually the random moves disappeared and a more informed set of moves began to emerge, moves that the program had learned through experiment seemed to boost its score.
It’s worth watching the supplementary video the DeepMind team included in the paper they eventually wrote. It shows the program learning to play Breakout. At first you see it randomly moving the paddle back and forward to see what will happen. Then, when the ball finally hits the paddle and bounces back and hits a brick and the score goes up, the program starts to rewrite itself. If the pixels of the ball and the pixels of the paddle connect that seems to be a good thing. After 400 game plays it’s doing really well, getting the paddle to continually bat the ball back and forward.
The shock for me came when you see what it discovered after 600 games. It found our hack! I’m not sure how many games it took us as kids to find this trick, but judging by the amount of time I wasted with my friend it could well have been more. But there it is. The program manipulated the paddle to tunnel its way up the sides, such that the ball would be stuck in the gap between the top of the wall and the top of the screen. At this point the score goes up very fast without the computer’s having to do very much. I remember my friend and I high-fiving when we’d discovered this trick. The machine felt nothing.
By 2014, four years after the creation of DeepMind, the program had learned how to outperform humans on twenty-nine of the forty-nine Atari games it had been exposed to. The paper the team submitted to Nature detailing their achievement was published in early 2015. To be published in Nature is one of the highlights of a scientist’s career. But their paper achieved the even greater accolade of being featured as the cover story of the whole issue. The journal recognised that this was a huge moment for artificial intelligence.
It has to be reiterated what an amazing feat of programming this was. From just the raw data of the state of the pixels and the changing score, the program had changed itself from randomly moving the paddle of Breakout back and forth to learning that tunnelling the sides of the wall would win
