TITLE: Computer Programs Have Much to Learn, and Much to Teach Us AUTHOR: Eugene Wallingford DATE: May 21, 2017 10:07 AM DESC: ----- BODY: In his recent interview with Tyler Cowen, Garry Kasparov talks about AI, chess, politics, and the future of creativity. In one of the more intriguing passages, he explains that building databases for chess endgames has demonstrated how little we understand about the game and offers insight into how we know that chess-playing computer programs -- now so far beyond humans that even the world champion can only score occasionally against commodity programs -- still have a long way to improve. He gives as an example a particular position with a king, two rooks, a knight on one side versus a king and two rooks on the other. Through the retrograde analysis used to construct endgame databases, we know that, with ideal play by both sides, the stronger side can force checkmate in 490 moves. Yes, 490. Kasparov says:

Now, I can tell you that -- even being a very decent player -- for the first 400 moves, I could hardly understand why these pieces moved around like a dance. It's endless dance around the board. You don't see any pattern, trust me. No pattern, because they move from one side to another.

At certain points I saw, "Oh, but white's position has deteriorated. It was better 50 moves before." The question is -- and this is a big question -- if there are certain positions in these endgames, like seven-piece endgames, that take, by the best play of both sides, 500 moves to win the game, what does it tell us about the quality of the game that we play, which is an average 50 moves? [...]

Maybe with machines, we can actually move our knowledge much further, and we can understand how to play decent games at much greater lengths.

But there's more. Do chess-playing computer programs, so much superior to even the best human players, understand these endgames either? I don't mean "understand" in the human sense, but only in the sense of being able to play games of that quality. Kasparov moves on to his analysis of games between the best programs:

I think you can confirm my observations that there's something strange in these games. First of all, they are longer, of course. They are much longer because machines don't make the same mistakes [we do] so they could play 70, 80 moves, 100 moves. [That is] way, way below what we expect from perfect chess.

That tells us that [the] machines are not perfect. Most of those games are decided by one of the machines suddenly. Can I call it losing patience? Because you're in a position that is roughly even. [...] The pieces are all over, and then suddenly one machine makes a, you may call, human mistake. Suddenly it loses patience, and it tries to break up without a good reason behind it.

That also tells us [...] that machines also have, you may call it, psychology, the pattern and the decision-making. If you understand this pattern, we can make certain predictions.

Kasparov is heartened by this, and it's part of the reason that he is not as pessimistic about the near-term prospects of AI as some well-known scientists and engineers are. Even with so-called deep learning, our programs are only beginning to scratch the surface of complexity in the universe. There is no particular reason to think that the opaque systems evolved to drive our cars and fly our drones will be any more perfect in their domains than our game-playing programs, and we have strong evidence from the domain of games that programs are still far from perfect. On a more optimistic note, advances in AI give us an opportunity to use programs to help us understand the world better and to improve our own judgment. Kasparov sees this in chess, in the big gaps between the best human play, the best computer play, and perfect play in even relatively simple positions; I wrote wistfully about this last year, prompted by AlphaGo's breakthrough. But the opportunity is much more valuable when we move beyond playing games, as Cowen alluded in an aside during Kasparov's explanation: Imagine how bad our politics will look in comparison to computer programs that do it well! We have much to learn. As always, this episode of Conversations with Tyler was interesting and evocative throughout. If you are a chess player, there is an special bonus. The transcript includes a pointer to Kasparov's Immortal Game against Veselin Topalov at Wijk aan Zee in 1999, along with a discussion of some of Kasparov's thoughts on the game beginning with the pivotal move 24. Rxd4. This game, an object of uncommon beauty, will stand as an eternal reminder why, even in the face of advancing AI, it will always matter that people play and compete and create. ~~~~ If you enjoyed this entry, you might also like Old Dreams Live On. It looks more foresightful now that AlphaGo has arrived. -----