Tuesday
Nov022010
Kasparov on Technological Innovation
Tuesday, November 2, 2010 at 11:06PM
Here's a summary of a talk Garry Kasparov gave for Palantir Technologies. One interesting remark comes in the context of computer chess, where he wishes that the triumph of chess programming had come from an artificial intelligence breakthrough rather than by brute force. What's interesting is that this echoes the approach taken by his great teacher Mikhail Botvinnik. Botvinnik was not only a world chess champion, but an engineer and programmer whose desire was just what Kasparov wished for: a truly AI-based chess program.
HT: Brian Karen
tagged Kasparov, computer chess, computers
Reader Comments (3)
He's visiting Google tomorrow, where I work, and I've been extremely excited for the last week. I suppose it will be something very similar to this talk today.
I realize he was specifically discussing his experience with Deep Blue, but doesn't Rybka and a few others use a pruning mechanism?
I also know that the quantum computers that are now just lab experiments will be the "quantum leap" [pun intended] he is talking about.
@Larry L: Modern engines are still doing basically what Deep Blue did. Pruning is just a heuristic that improves performance at the margins without fundamentally altering the approach. (Deep Blue probably employed pruning in some fashion, as the idea has been around a while, but I don’t think DB’s algorithm has ever been released, so we don’t know for sure.)
In the terminology of Claude Shannon, who invented computer chess as we now know it, practically all of the useful engines that have ever been developed employ “Type A” (brute force) algorithms. For the most part, they are getting better mainly because the hardware is getting faster. Kasparov sounds like he is longing for a “Type B” breakthrough, a computer that would think about chess the way humans do.
To give an example, if you open with 1.e4 and take away the computer's opening book, it will analyze all 20 of the legal responses as if they were equally likely to be the best move. Obviously, the evaluation function quickly concludes that 1...e5 is a better response than 1...a6 (it controls the center, opens developing lines, etc.). But it’s still giving equal attention to all 20 before reaching that conclusion.
Humans do not do that. A human does not devote equal attention to every legal move in a given position. Kasparov appears to be saying that if someone could actually make a computer play strong chess while “thinking” about positions the way humans do, he believes that breakthrough would be usable in other fields. He is probably right about that, but we will never know until someone actually figures out how to do it.
I believe there have been some attempts at developing “Type B” chess engines, but they have never been nearly as good as the “Type A” ones. Modern engines do, of course, have some “chess knowledge” abstractions built into them, especially in endgames, where a true brute-force approach could lead to some absurd moves. But those rules are built as “exceptions” to what is, at its root, a “calculate-everything” approach.