John Scanlon wrote: > Is anyone interested in discussing the use of formal logic as the > foundation for knowledge representation schemes for AI? It's a common > approach, but I think it's the wrong path. Even if you add > probability or fuzzy logic, it's still insufficient for true intelligence. > > ... > Any thoughts or arguments? Or am I just repeating something everyone > already knows? (I honestly don't know.) > > J.P. Thoughts or arguments? OK, here's a few tho I tend to think of the mind as divided into four functions which are represented thus: 1: Air: formal systems, rule based systems: True/False 2: Earth: model based systems: Useful/Useless 3: Fire: ranking based systems: Good/Bad 4: Water: goal oriented systems: ?? My "model" is that everyone starts from some base, and builds the other parts on top of it. There are some limits on which can be built on top of what that reduces the number of possibilities (See C.G.Jung) to around 12, so that one may get a mind shaped like: 1:2:4:3 (Note that when the first three are chosen, the last is determined, as all four functions are always present)
Naturally there's more to it than this, but note that logic is only ONE of the components, and all of the components are essential. It's not clear that each of the functions uses the same kind of data representation. (It *is* clear that they value data differently.) But consider, bot air and fire could (apparently) use a singly linked list as a data structure, while earth probably couldn't. Earth and Water appear to be more process oriented than are Air and Fire. Fire is quick, and air is slow...i.e., Fire orders things into a ranking, and takes the top one first, while air appears to be an expert system. With fire the processing is done "off-line". Air will use lemmas and theorems previously proven, but each new proof is still slow. Earth appears to operate by constructing simulations (normally annotated sensory imagery, but kinesthetic, etc. can also be included). Water...I have less insight into how water operates. I would guess that it constructs models of potential goal states, and then applies alpha-beta pruning to select between them. Note the time variation in the different approaches: Fire is fast, but error prone. Air is slow, but certain. Earth ... the speed depends on the amount of abstraction that has gone into the model, but it's always slow compared to Fire, and can be slow compared even to Air...and it's never as certain as Air can be, but it can reach answers that are beyond those that can be reached by Air. Water is continuous. It never reaches completion. (Should never reach completion?) Water is future oriented in a way that the others aren't, always searching for what "should" be the next goal. A note on the nomenclature: I derived this nomenclature from C.G.Jung, who got it from alchemy. I've mutated it dramatically for my own ends, but I find it much more amenable to thought than "logic, emotion, intuition, and practical sense" which would be the "ordinary" words closest in meaning to what I mean. If I used those terms *I* would get snarled in the various meanings that they carry, much less anyone I might talk to, and I work with numbers carry an inherent sense that some particular one is prior to some particular other, so indexing is also a suboptimal choice. The only other choice would be to invent terms from scratch...Latin and Greek have been too heavily mined in this field to be a good source. That's a normal approach, and often good, but to mean the "4 elements" is no less sensible. P.S.: You don't know that your model of mind is complete and correct until you've implemented it. Then you need to run the test. Up until that point you are working with metaphysics and philosophy...not bad, but not practical. Once you implement it, it is practical. Then you have success, failure, or a chance to improve things. And it is vital to avoid certain modes of failure (or should I call that success?). Also note that there is the crucial omission of the axioms of the system, the instincts. How does it chose it's goals? At the primal level, what is Good and what is Bad? These questions need to be answered before the mind is built. At this point it not only doesn't know what a person is, it doesn't know what a physical world is. Most people on this list seem to be heavily involved in Bayesian probability logic. Personally, I think people have more of a tendency to use pattern recognition, but the two techniques are probably just different roads to the same end. (And not completely separate, either!) My suspicion is that it will turn out that Bayesian probability is more suited to processing with only a few threads, where pattern recognition is more parallelizable, but this is just a guess on my part. Few are devoting much time to logic, but that's probably because expert systems have pretty much solved that piece. I'm sure that there will be improvements, but it can be handled already at a "better than a human" level. Interfacing expert systems to the other pieces, however, may be more challenging, and designing the other pieces is also quite challenging. (My first try came to a few quick blockers, and I've been trying to think how to try next.) ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
