Why is the Google Deep Mind win a big deal?
Google’s AI program AlphaGo beat the best Go player in the world Lee Se-dol (just the first game in a best of 5 series as of now, I’m pretty confident it will win overall). This is almost 20 years after IBM’s Deep Blue beat Kasparov (the best chess player at the time). So why did it take so long and why is it a big deal?
20 years ago computers, more precisely artificial intelligence couldn’t do much except win at Chess, fly planes, control robots in factories and make basic weather predictions. Why Chess and not Go? Chess and Go with similar board sizes, black and white pieces and turn based adversarial game play seem quite similar. That’s where the similarity ends. Chess and Go for computers are very very different. In Chess there are on an average 35 moves possible for a player. In Go, that number is a whopping 250, after 40 moves the difference becomes ~10³⁴ in terms of added complexity.
The precise reason why a computer could win at Chess was because it is possible to calculate the relative strength and weakness of a position just by looking at the board. Each piece on the board has a value, the more pieces you have the more likely you win. A queen is worth 9 points a rook worth 5 and so on. Of course the position matters but it is relatively easy to calculate the best outcome in terms of number of points 5 moves later. The scores for pieces can be very easily calculated by looking at a few 100 games.
This is where Go is a very tricky game. Looking at the board and saying who is winning is very very difficult, especially in the early stages of the game. Even at the end of the game when Lee Se-dol resigned it wasn’t obvious why. Go has been described frequently as a Game where intuition is more important than calculation. Intuition in this case can to some extent be summarized as the strength of a given position which is difficult to numerically calculate.
Take for example this picture, is it possible to intuit what you should do next to avoid an accident? Anything you do now might cause an accident 3 seconds later. All the other cars are moving and everything 3 seconds later is going to be different. This is Go. Whereas Chess can be thought of more like calculating the shortest path home with roads closed at random intervals. You constantly know how far you are from home. You might find a road closed, but you probably won’t need to go all that far to avoid the closed road.
Humans are pretty good at intuiting long term value with limited information. Entrepreneurs know how their companies are doing even if they don’t get a valuation done daily. A painter can paint without forming the entire image in their head, she understands that if a stroke is applied nothing else can go there. Most humans with some practise can drive cars and avoid accidents.
The power of intuition is what seems to have taken 20 years to build for computers. Don’t worry, computers aren’t rising against us tomorrow. All this means is that AlphaGo was able to do is assign a numerical value to a position. The equivalent of figuring out the value of each piece in chess.
Why is this important? Most problems that computers will need to solve in the modern world are optimization problems where the value is not known at the end of each decision. Google didn’t just understand how to solve the problem for Go, it figured out how to apply this methodology to any problem.
Computers were bad at solving such problems and now seem to have got really good because of Deep Learning. Deep Learning is just a fancy term for being able to “understand” and “learn from” large amounts of data. Understand and learn from here are just words used to explain the math that goes into storing data in a format a computer can quickly look at when it needs to take a decision. This is the computer’s way of making a cheat sheet for itself.
If I slow down right now, do I reduce the chances of hitting the old lady and also avoiding the car behind me? Seems pretty obvious doesn’t it. It only took a computer 20 years to learn this. The fun/unfortunate thing about Moore’s Law is by next year no human will even come close to competing with AlphaGo.
At least we think we are better drivers for the time being.
