Friday, May 12, 2006

Happy Deep Blue Day

In 1997, Deep Blue became the first computer system to defeat a reigning world champion in a match under standard chess tournament time controls. You can read about it at
http://en.wikipedia.org/wiki/Deep_Blue

I like to celebrate this day every year by wearing blue, drinking milk, and talking about Deep Blue. It's an interesting case in AI because it's a great example of how AI is perceived as a fanishing science-- that is, as soon as a computer program can do something, that something is no longer perceived as demonstrating intelligence. As a result, AI researchers wrying joke that AI stands for "Almost Implemented."

There are three reasons people perceive AI as a failure. The first is that the goal posts keep moving. There was a time that calculation, and doing lots of arithmetic, was perceived as very intelligent behavior. Now that nobody blinks when their watch can do calculations faster than any human being, it's hard for people (AI researchers included) to think of calculating as intelligent. The goal posts moved. Whoops. We'll have to kick farther. This is true also of chess. Naysayers used to say a computer could never play grandmaster-level chess because it required real intelligence. Well, in 1997, we had an objective outcome. A computer program could indeed beat a grandmaster. Well, the naysayers were quick to change their tune, just as they always do: Now playing chess is no longer an act that requires intelligence. Whether or not there is validity to this I will adderss below.

The second reason is related to the first. People have a mystical view of intelligence, and once we fully understand how a computer is doing something, it fails to have mystery, so it does not qualify as intelligence anymore-- it's just search, or it's just cranking through an analogy. How do people do the same things? We don't know. Maybe they are doing heuristic search sometimes-- we just don't know! Intelligence will not be de-mystified until we better understand how people think. When we can reduce human thought to "mundane" algorithms like search or connectionism, maybe the naysayers will grumble some acceptance of artificial real intelligence.

The third reason is that we were too optimistic in the 1950s about how easy AI would be. We made bad predictions that I think people are only just starting to forgive us for. I won't go into this reason more because it's not important for this essay, but yes, some of the blame falls on us.

So, Deep Blue. Intelligent or not?
Although Deep Blue had a great deal of chess knowledge, its main power comes from algorithms created long ago in AI. The reason, some will say, that Deep Blue won was because it had enough computing power behind it, not because it was intelligent. As though any fast computer could beat Kasparov!

Here's another way to look at it: The algorithms we came up with long ago were good-- they were the right ones. It's just that they were ahead of their time. The computers of the time were not powerful enough to really show off what they could do. But since we came up with the algorithms long ago, we no longer get credit for them. Is that fair? (hint: "no.")

Another objection to Deep Blue is that it considers many more moves than Kasparov. It's more of a brute force search, so people want to discount it for this reason. It seems pretty certain that it uses different algorithms than does, say, Kasparov, but is this a good reason to say it's not intelligent?

Intelligence is one of those weird words. Do airplanes fly? Yes, they do, even though they don't flap their wings. Do boats swim? No, they don't. Not in English anyway. In Russian, boats do indeed swim. Similarly, intelligence is one of those words for which people disagree on whether a system that does an intelligent act differently than a human should be considered intelligent. Do we really want out opinions on whether or not computers can think to be dependent on the nature of people's commonsense notions of what intelligence means? In English?? (hint: "no.")

So the question of whether or not Deep Blue is actually intelligent is, at present, not an empirical question. Someday it might be.

So acknowledge this, have a drink of milk, wear blue, and celebrate Deep Blue day. Whether or not you want to call AIs intelligent, they are landing our airplanes, optimizing prices, giving you recommendations on amazon.com, understanding our voices on the phone, checking to see if people are forging our checks, teaching us about chess end games, and beating our chess masters. Cheers.

2 comments:

Christy McGuire said...

OK, I give up: what's milk got to do with this?

Jim Davies said...

I love milk but by digestive tract does not. I'm always looking for excuses to drink it.