At this point in time, I'm for less gedankenexperiment and more coding. Cheers, ~PM
Date: Mon, 19 May 2014 15:51:44 -0400 Subject: Re: [agi] The Parts Knowledge Can be Used to Make Many Generalizations From: [email protected] To: [email protected]; [email protected] If George is right his program will (eventually) work.Refining this you get a slightly better statement. If and only if George is right will his program eventually work.But which is the better statement? They both have some merit. It is unlikely that George will get his program to work even if he is wrong, but it is a possibility. So both statements have some value. In fact there are quite a few useful statements that could touch on those conceptual objects. Now look at this:If Jim is right George will not be able to get his project to work. Now suppose that some time goes by and George, against his own judgment, figures out what Jim is saying and he begins incorporating some of the necessary sense of reality that would make his project feasible (according to Jim). Now all the listed logical statements previously made are irrelevant. They can't be validated - even if George gets his project to work - because the presumptions changed slightly. You can attempt to assign 'blame' for the failure of the logical statements to retain their relevance. The problem was that new information was introduced to the system and that new information made the older logical statements irrelevant. While understanding why a theory failed is an important part of learning, it does not matter why the theories failed to sustain their description-of-reality value to invalidate the use of elaborate logical systems of thought. All that is important is the recognition that they do not even have local consistency unless you are defining local consistency as the assertion of the application of the logical system to the world case. Once an application error becomes apparent, or as in this example, comes to pass, the validity of the logical system is only useful if it might one day be used to analyze some other situation. So the sense that the methods would be locally consistent even when they aren't globally consistent isn't really reasonable - unless you are defining local systems as an assertion of the premises and the premises of the application of the system. Jim BromerJim Bromer On Mon, May 19, 2014 at 3:13 PM, Jim Bromer <[email protected]> wrote: OK, that's cool but the logical framework that Michalski is talking about is a representation system but not a true logical system. It can be used to represent some interesting relationships of thought-stuff. Jim Bromer On Mon, May 19, 2014 at 3:00 PM, Piaget Modeler via AGI <[email protected]> wrote: The forward and backward confidence parameters are adjusted by the type of knowledge transmutation performed over the knowledge base. My opinion is that Global Consistency is not important for an AGI system. Cyc handles consistency by using microtheories, or collections of propositions and inference rules. Each microtheory is consistent, but if taken altogether, there will be global inconsistencies across microtheories. In PAM-P2 we take a similar approach. We have viewpoints which are similar to Lenat's microtheories,but we also don't really care if premises are inconsistent. We embrace inconsistency and rely more on activation to sort things out. (PAM-P2 is still in process so we'll let you know how things turn out, and whether or not we modify our position on this point.) But I think Michalski's introduction of merit parameters and probability into his logical framework has merit, no pun intended. ~PM Date: Mon, 19 May 2014 14:42:31 -0400 Subject: Re: [agi] The Parts Knowledge Can be Used to Make Many Generalizations From: [email protected] To: [email protected] But what was the basis for the forward and backward confidence? The problem is that this is still a logically inconsistent system posing as a logically consistent system. I can't create logically consistent AGI systems, but maybe I am just more honest about it. The consequence of this is that his logical system is merely a representational system. I've known guys who tried to talk about ideas and then thought they could emphasize them with pseudo-formalization (or maybe partial-formalization). Nothing wrong with that - unless they thought that they were actually formalizing their various conjectures. But they were only simplifying the representation of very narrow ideas by using formal symbols and stuff. So the formalization for these kinds of things are not truly consistent abstract systems that can be used clearly as the programmatic basis's for computer programs. It is a notation for an informal system that has limited applications. Nothing wrong with that, but let's be honest about it. Jim Bromer On Mon, May 19, 2014 at 1:09 PM, Piaget Modeler <[email protected]> wrote: Michalski injected probability into his system with the notion of merit parameters, for forward and backward confidence in statements, implying that a purely logical system might be insufficient to handle real world phenomena. ~PM. Date: Mon, 19 May 2014 12:25:22 -0400 Subject: Re: [agi] The Parts Knowledge Can be Used to Make Many Generalizations From: [email protected] To: [email protected]; [email protected] Since false assertions can be mixed in with good assertions, the potential complexity of an idea (a reference) cannot be neatly or easily categorized. Michalski did mention that some logical relations are truth-preserving and some are not but the whole idea of an underlying logical system is that some important relations may be derived based on the abstractions. (Just as new mathematical theories are discovered.) The important abstract relations would typically be discovered by a close study of the applications of these ideas to real world situations (or to the situations that the mind can consider). But since references will contain hidden combinations of other references and since false assertions will tend to be embedded along with good assertions and since the reasons that would support the insights would also be based on similar combinations of information, my conclusion is that the potential benefit that the elaborated logical system might provide may well be compromised and even fatally flawed by inappropriate assertions and assumptions. So while I would use logic in arbitrarily constrained systems, I feel strongly that the underlying 'logic' of an AGI system has to be comprised of the description of the construction of the relationships of the references. In other words it is a dynamic descriptive system that must tend to limit the assumption that the systems are based on broad underlying generalizations. The generalizations that I have in mind will tend to be specialized (even though I do suppose that similar methods can be used with them when the methods are fit to the application through trial and error.) I really don't have a solid idea what verification will consist of, but I am supposing that systems of insight that can lead to reliable interactions will have some value. Jim Bromer On Mon, May 19, 2014 at 8:59 AM, Jim Bromer <[email protected]> wrote: Thanks for the reference to Inferential Theories of Learning. I found something on the Internet. http://www.mli.gmu.edu/papers/91-95/MSL4-ITL.pdf I am glad to see that someone has been interested in looking at learning as the ability to see how different kinds of inferences may lead to useful knowledge. I have written (in these groups) about how I believe that conceptual projection and the integration of different kinds of knowledge is very important to AGI. So these can reasonably be considered as different kinds of inferences similar to Michalski's definition. My feeling is that an emphasis of the formal - or general - processes that the author likes to rely on may be a misrepresentation error. Some of his ideas are good, and the examples are interesting. However, in detailing some fundamental abstractions (programming abstractions) he is in effect declaring these as special fundamental abstraction-to-generalization methods. Maybe I should say it is a fundamental attribution error. The problem is that the combination will certainly, and the individual application will probably lead to contradictions of the theory. In order to avoid this one would have to create fundamental application definitions which assert the kind of rule that is being applied to an actual problem. In other words, the attempt to rely on a fundamental abstraction or general rule won't work. I realize that Michalski is aware of this, at least at some level, but in his assertion that there is some kind of competency test, (I forget what the test was based on) he is implying that false assertions can be eliminated. They can't be. Sure, I will be using some kind of logic in my model. But, the underlying principles in my model does not consist of an abstraction of logic but simply an abstraction of construction that will describe, to some extent, how the relations of a concept were formed. Jim Bromer On Sun, May 18, 2014 at 1:15 PM, Piaget Modeler via AGI <[email protected]> wrote: You may want to read The Inferential Theory of Learning by Ryszard Michalski. He and Gheorghe Tecuci of GMU did some very good work in Reasoning. It may be helpful in your thinking about this topic. ~PM Date: Sun, 18 May 2014 12:51:40 -0400 Subject: [agi] The Parts Knowledge Can be Used to Make Many Generalizations From: [email protected] To: [email protected] In order to make detailed insights feasible, they need to be generalized. I bet that almost everyone who will read this in 2014 will misunderstand what I meant at first. I don't mean that many pieces of knowledge should be generalized into one idea, but that the parts of many individual pieces of knowledge can be generalized into many individualized generalizations. I am sure that this is being implemented in some nlp, but only at a very rudimentary level. The possible abstractions and combinations are uncountable. This process then would have the capacity for immense individualization. But it is not as simple as it might seem because computer programs that can keep track of, refer to and wisely use an immense number of possible combinations are not simple. Jim Bromer AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
