Does your intuition tell you that computers can’t develop intuition? If so, then you might want to reconsider in light of this week’s news.
Computers have been better at chess since Deep Blue beat Garry Kasparov in 1997, but this week Google’s AlphaZero, a general-purpose artificial intelligence (AI) absolutely crushed the reigning computer champion, winning 28 games out of 100 and drawing the rest (no losses at all). It did not win because it computed faster or with a better brute-force algorithm. It won by having better chess intuition than any human. And it developed this intuition all on its own.
AlphaZero was programmed only with the basic rules of chess: these are the legal moves and you win by checkmating the opponent’s king. It was given no help at all with strategy. It was not even told that a Queen is a better piece than a Pawn, much less was it programmed with any hints about how to develop an intuitive feel for a good position.
On the other hand, its opponent, a program called Stockfish, was programmed to know not only the rules but dozens of heuristics with which to evaluate any given position. For example, its human programmers told it that Knight is worth roughly three Pawns, that a Bishop is most powerful when it can look down a long diagonal, etc.. These heuristics have been developed and fine-tuned based on centuries of human experience with the game. We could say they are human intuition put in the form of a computer program.
So how did AlphaZero learn enough strategy to beat the best computer program in the world (which, in turn, was already much better than any human)? It played games against itself. Yes, it was entirely self-taught.
How long did it take for AlphaZero to learn all this strategy and hone its intuitive sense?
Yes, in just four hours an artificial intelligence, given nothing more than the rules of the game, was able to vastly surpass the knowledge that some of the smartest people on the planet had developed in the half-millennium since the modern version of the game was invented.
A chess program is very good at following a chain of “if I do this and he does that…”. However, it’s not possible for even a computer to follow every possibility; there are just too many. It must know when its calculations have reached a position that is not worth analyzing further because it is clearly very good or very bad. In addition, when it reaches the end of a sequence of “if I do this and he does that…”, it must evaluate the resulting position based on the sorts of heuristics I mentioned: a Bishop on a long diagonal is good, and so on. We can call this evaluation a chess player’s intuition — his or her “feel” for the position without doing any further computation.
What’s most striking about AlphaZero’s victory is that it actually evaluates moves nearly 1,000 times slower than Stockfish but more than makes up for this by doing a better job of evaluating the positions that those moves reach. That is, it has a better intuitive sense. And again, it developed this intuition all on its own. In four hours.
Within my lifetime, intuition was considered something that computers could never develop. Yet here we are.
Where do we go from here? Well, if you make a good living by giving advice based on your mastery of a large body of knowledge (law, medicine, and engineering come immediately to mind), you should enjoy your job because it might not be there in a few years. Or to look on the bright side, the services you now care deeply about will soon be rendered with much more excellence and with much less cost.
I guess that’s good news, right?