Friday, October 27, 2006

Constructing HAL

An interesting presentation by Doug Lenat, AI pioneer and founder of cycorp, at google.



Cyc plans to feed all the information in the world into a computer, and then let the computer use first order logic to reason with the information to answer questions. Now before we go all 2001 over this, its interesting that Lenat has been working at it for over twenty years, and the system can now answer questions like 'Is President Bush inhaling or exhaling right now?' correctly, with an 'I don't know'.

To be fair, some of the examples in the presentation are impressive, but is it a victory for AI, or for Moore's law? I am not a great fan of the Deep Blue approach to AI problems. It seems the way to solve an AI problem is to destroy it.

On the other hand, Lenat does point out a very serious problem with gradual learn-by-experience approaches: a Catch-22 situation where the less the system already knows, the harder it is for it to learn further or evolve. So classical AI might be a good bootstrap to give learning systems a headstart. That is actually a very good idea, and I haven't seen anyone doing that.

4 comments:

Anonymous said...

Artificial Intelligence: A Philosophical Introduction by Jack Copeland discusses Doug Lenat's work.
May be it is more from the perspective of whether what human's do is anywhere close to how Doug Lenot is doing it, than from whether it will solve the problem at hand. Still gives a handle on the issue.

sbkt said...

Thanks for the tip...anonymous ;) .

sbkt said...

Some more comments...
I think Lenat once said: 'Intelligence is 10 million rules'. I do not disagree with him on that. If we are not talking Einstein-type intuitive insights, most day-to-day activity is just a lot of remembered rules. But the hard part is not storing all these rules, but knowing exactly which rules apply to a situation.Rule-based systems usually retrieve pretty much every rule in their system and try them all, till they find one which works.Suppose you have a million rules. Where do you begin this process?To my knowledge,Lenat hasn't focused on this.He used predicate calculus with forward-chaining.As you know,that is extremely expensive.
Also what he is doing is v far from elegant. It is the sledgehammer approach IBM used with Deep Blue.I get irritated when people say the chess problem is solved, cos it is far from solved. It has just been killed. A chess pro like u (;) ) should know players don't try every possible move till they find the right one.That is what Deep Blue does, and that is a victory for Intel that they can do that, not IBM.

Anonymous said...

I am not clear what you mean by rules here, but if you mean sequential processing of the 'if-then-else' kind, it seems doubtful that most of our day to day activities are done that way. Riding a bike does not seem rule-based. You'd know better, but how about how we do pattern matching and face recognition?

However you win in chess, if you do, you have solved the problem. There is only so much elegance, as you see it, should count for. A rule-based approach may work well for things like chess, where it is easy to parameterize, to identify rules and to have an ontology.

The problem with rule-based approaches is not that they kill a problem and not solve it. It is that they can probably only work for specific types of problems.