On 10/8/07, Edward W. Porter <[EMAIL PROTECTED]> wrote: > > --(1) How are episodes represented in NARS?
As "events" --- see http://nars.wang.googlepages.com/wang.roadmap.pdf , pages 7-8 > --(2) How are complex pattern and sets of patterns with many interrelated > elements represented in NARS? (I.e., how would NARS represents an auto > mechanic's understanding of automobiles? Would it be in terms of many > thousands of sentences containing relational inheritance statements such as > those shown on page 197 of "A Logic of Categorization"?) Not necessarily "inheritance statements", but "Narsese statements" in general. > --(3) How are time and temporal patterns represented? As "events" or "operations" --- again, see http://nars.wang.googlepages.com/wang.roadmap.pdf , pages 7-8 > --(4) How are specific mappings between the elements of a pattern and what > they map to represented in NARS? As various types of "relation", which are special type of "term". > --(5) How does NARS learn behaviors? Mainly through "procedural reasoning" --- the above paper has a brief description, and the book has a more detailed description, though I'm still working on the details. > --(6) Finally, this is a much larger question. Is it really optimal to > limit your representational scheme to a language in which all sentences are > based on the inheritance relation? Well, it indeed deserves a longer answer. First, NARS doesn't use the inheritance relation for all sentences --- in the current implementation, there are four relations in the memory: inheritance, similarity, implication, and equivalence. Though the later three are derived from inheritance conceptually, they are processed on their own. Second, to say "the memory contains four basic relation types" doesn't prevent the system from representing and processing other "user-defined" relations --- see the above paper, page 5, "Products and Images". It is just that only the four basic types have "fixed meaning", while the meaning of the other relations are learned from experience. > With regard to Question (6): > > Categorization is essential. I don't question that. I believe the pattern > is the essential source of intelligence. It is essential to implication and > reasoning from experiences. NARS's categorization relates to patterns and > relationships between patterns. It patterns are represented in a > generalization hierarchy (where a property or set of properties can be > viewed as a generalization), with a higher level pattern (i.e., category) > being able to represent different species of itself in the different > contexts where those different species are appropriate, thus, helping to > solve two of the major problems in AI, that of non-literal matching and > context appropriateness. > > All this is well and good. But without having had a chance to fully > consider the subject it seems to me that there might be other aspects of > reality and representation that -- even if they might all be reducible to > representation in terms of categorization -- could perhaps be more easily > thought of by us poor humans in terms of concepts other than categorization. NARS doesn't rule out problem-specific and domain-specific representation, though they are handled at a different level. Narsese is like the "native language" of NARS, though based on it the system can learn various types of "second/foreign languages" (including natural languages). However, this is different from merging those languages into Narsese. See the above paper, pages 9-10, "Natural languages", for a brief explanation. > For example, Novamente bases its inference and much of its learning on PTL, > Probabilistic Term Logic, which is based on inheritance relations, much as > is NARS. But both of Ben's articles on Novamente spend a lot of time > describing things in terms like "hypergraph", "maps", "attractors", "logical > unification", "PredicateNodes", "genetic programming", and "associative > links". Yes, perhaps all these things could be thought of as categories, > inheritance statements, and things derived from them of the type described > in you paper "A Logic of Catagorization", and such thoughts might provide > valuable insights, but is that the most efficient way for us mortals to > think of them and for a machine to represent them. NARS and Novamente surely still have some "family resemblance" left --- for a "family story", read http://www.goertzel.org/benzine/WakingUpFromTheEconomyOfDreams.htm These two systems have many similarities, as well as important differences, on which I and Ben have debated for years. It is too big a topic to be addressed here. Pei ----- 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/?member_id=8660244&id_secret=51308317-fadbbc
