On Sunday 11 March 2007 15:07, YKY (Yan King Yin) wrote: > My main point is: a unified KR allows people to *work together*.
That would certainly be nice, but I have yet to be convinced that it's possible :-) > Let's look at the alternative, which is even more dismal: you have many > representations and each is governed by its own logic and has its own > inference algorithms. There'd be a huge problem of interoperability and > communication. Certainly. Yet I am convinced that that's how it actually works. Someone who came from a theoretical pure communist economy where there was only one organization that everybody worked for, would be aghast at the random madhouse of a market economy. But the market not only works, it's enormously more adaptable (in AI terms: less brittle). > Are you suggesting to have separate modules for vision, speech, > naive physics, arithmetic, social reasoning, etc? In common-sense > reasoning, steps from all these areas need to be chained together. I'm suggesting we have separate modules for turning the pages of hardcover versus paperback books, separate modules for walking on sand versus gravel, separate modules for reading Roman versus Helvetica type. I think commonsense "reasoning" is an illusion (if it's even that -- nobody really has the sensation of reasoning out something that's "common sense." We just seem to know it effortlessly). I think it's mostly memory -- a customary supply chain gets built up in the market of mind. I think that *after the fact* something like EBL happens. > Perhaps there could be *variations* within each module, but they should > share an overall unified representation, IMO. Sure, at some level -- any one neuron looks like any other at a chemical or voltage level. But I think the level where all the modules are the same is below the level where one would tend to call it a "representation." I'm working with an inter-module representation of numeric vectors -- but the interpretation of a vector varies with each (sender, receiver) module pair, so there are possibly n^2, rather than merely n, different representations (but fewer in practice :-) > Let's see how YOU could work with a million representations? Why make > it more difficult for ourselves? But I DO work with a million different representations in my head -- that's my point. The bottom line is that you don't have to deal with them explicitly if the system is capable of forming new ones and integrating them with the rest automatically, and if it isn't, it isn't going to work anyway. > This is my philosophy: we should allow enough FLEXIBILITY such that the > system is robust enough, but not more than that. For example, the Wright > brothers designed wings that are flexible *enough* for controlled flight, > but they did not go all the way to build BIRD'S WINGS. The Flyer couldn't land and take off from tree branches either, or fuel itself by eating insects. Even so, I would agree to the extent that I bet we can take three or more orders of magnitude complexity off the brain and yet build a working thinking machine. But it'll still have a million representations. > The same prioritization will be present in the logical system as well -- > not all beliefs are of the same ranking in a truth maintenance system -- > although it may not display exactly the same Wason card phenomenon, but we > will certainly design heuristics for efficient reasoning within the logic > framework, especially to deal with frequent / critical situations. ... > Yes, there are local *rules* that govern various domains; eg rules that > govern natural language are different from rules that govern vision. In > fact the locality of rules can be exploited to make rule-lookup faster. > But the overall representation can be very uniform. Again, sure, everything can be bits, or words, or even symbolic clauses. But once you've done that you've basically just said "OK, our implementation language is going to be Prolog" and you still have almost everything yet to design. (a) It's absolutely true that an AI can be programmed in Prolog if it can be programmed at all. (b) Using logic as a substrate has been tried for 50 years and it doesn't seem to get you there any faster than any other decent programming language / representation. > The KR scheme itself does not contain a lot of knowledge; it is just a > knowledge *container*. The real stuff is represented as rules -- and these > rules can be *either* hand-coded or learned. That's one strenght of it. Oddly enough, I actually wrote a system that generates most of the clauses that it executes, not too long ago. It's an interpreter, in Prolog, for a slightly higher-level language that will in turn be used to write even higher-level stuff in the overall AI. But there are many levels separating that from anything you'd recognize as representing anything the AI actually knew. Josh ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
