One would definitely want to include encapsulation and abstraction as requirements of an AGI system.Otherwise it would be overwhelmed during search. ~PM.
Date: Thu, 6 Dec 2012 10:06:30 -0600 Subject: Re: [agi] Internal Representation From: [email protected] To: [email protected] I've pointed to the Encapsulation concept from Object Oriented paradigm before as a means for limiting complexity. Packing up highly interrelated subsystems recursively and extracting out summaries of important/relevant expected behavior for each system expressed independently of the details of the underlying processes that lead to that behavior, allows for the search space to be fragmented automatically based on degree of connectivity. The expected behavior of the system can be summarized by running a simulation-based search among the interactions of only its subsystems, given the events that might affect the system. It is a means for automatically generating effective heuristics which functions independently of the knowledge domain. A summary is less complex, by definition, and so working with summaries reduces the combinatorial explosion to something more manageable at the price of sometimes missing seemingly insignificant details. This also explains why we think in terms of objects in the first place. There is no reason to think that just because we draw a line around a set of phenomena and call them an object or a kind of object, that the instance or type we have isolated is necessarily an inherently distinct thing in reality. The bounds we draw around those phenomena (encapsulating them) are in our minds, not in reality, but they are difficult to recognize as such because they're the very things we use for recognition. Why would we add this new information to a system which is already overwhelmed by information? To organize that information flood into something manageable. On Thu, Dec 6, 2012 at 6:08 AM, Jim Bromer <[email protected]> wrote: Piaget Modeler <[email protected]> wrote: Integration is combining different concepts via their attributes akin to crossover or memetic recombination. ...I think complexity can be an issue if you have one huge problem and one huge model. Search spaces become ridiculous very fast. However, if you engineer a system so that you are not dealing with one large problem but lots of small problems, I think its feasible to engineer solutions to manage the complexity of each problem reasonably. ------------------------------- The wikipedia definition of memetics was interesting. Assuming that I can make a pretty good guess about how your idea of memetic recombination might work, I would say that your imagined usage of the method has some serious problems. First, a meme cannot be modelled in the same way a superficial data string can be so the comparison of memetic algorithms to recombination in genetic algorithms seems fanciful. Secondly the idea that the attributes of a concept might be clearly differentiated in an automated system that is able to learn and then used to clearly integrate different ideas seems unlikely. I do not think the concept is impossible, I think that it is complicated. It is a problem of complexity. You mentioned that you thought you can avoid complexity by using many small search problems. Although I cannot point to this or that study which can drive this point home, I do feel that there is ample evidence that domain restricted learning has not worked in AI just because we need to use concepts outside of the domain in order to understand those concepts which are strongly within the domain. (By the way, here is where an imagined efficiency of using weighted evaluations can really turn to nonsense. You can't eliminate the need to look outside the domain to determine meaning or relevance just by putting a numerical value on how much a meme belongs to a particular domain.) So to simplify this (so it does not sound like the product of a chat bot to you), a fully automated system cannot be expected to reliably determine the attributes of a concept (or meme) in order to use them to flawlessly integrate concepts. So assuming that your program would be able to achieve some viable integration one has to assume that the integration would be very rough. Your program might be able to make a crude haphazard matrix of relations between the concepts but how would it continue to refine this process without encountering some serious complexity? Here is where the hollow functional identity argument (as I called it) turns to delusion. 'If the program could do some crude thinking then it would be able to use the same processes to do more sophisticated thinking.' This is a philosophy that many AI programmers live by. Is it reality? We all know from personal experience and observation that there is a limit to intellectual ability and those limits have something to do with complexity. To believe that you can avoid complexity just by keeping the search space to a minimum doesn't sound insightful. Does that method work for you personally? Of course not. It is obvious that you have been trying to find methods for your program by broadening the search space of your studies. This is just the opposite of what you formally declared was a solution to complexity. It is an obvious contradiction. When you can't figure something out you sometimes expand the search area and you sometimes limit the search area. These method don't always work but they are important steps in learning and trying to find solutions. Jim Bromer On Wed, Dec 5, 2012 at 12:18 PM, Piaget Modeler <[email protected]> wrote: Jim, your prior e-mail reads like you are either a chatbot or are attempting NLP (Neuro-Linguistic Programming) or DHE. Just ask, my answer may be yes or no. My own reason for assisting is that I'd like you to understand my approach. Differentiation IS conditional branching. Observation is receiving sensory stimuli. Coordination means making inferences.Integration is combining different concepts via their attributes akin to crossover or memetic recombination. Please define verification. It may be what I call correlation. Cheers! Imagine, NLP via e-mail. Whooda thunk it? Date: Wed, 5 Dec 2012 07:54:16 -0500 Subject: Re: [agi] Internal Representation From: [email protected] To: [email protected] I agree with Piaget's remark. I am going to conduct an experiment. I want to see if I can get you to solve a problem for me. So I am going to keep track of our conversation by keeping notes on particular issues related to this experiment. It is unlikely that you would be able to solve a particular problem that is of interest to me, so I am going to be looking for an unexpected solution to some related problem that I will pick up somewhat serendipitously from our conversation. The best way to get you to cooperate with me on this is to get you talk about the thing you are interested in. However, the solutions to the problems of your projects probably will not be the solutions to the problems of my projects, so I have to find a way to get you talk about something that is common to both of our projects. So I have gotten you to describe some ways that your program can apply imagination to problem solving. Your seem to acknowledge that integration is a part of the process, but you haven't acknowledged that complexity is a problem. So now, in order to get you to continue discussing this I have to back off from talking about complexity and emphasize the problems of 'verifying' and integrating internal projections. I will review your message in response to my question of how your program will use imagination, and I will copy that response into my notes. Now that I have reviewed some of your previous messages I see that you mentioned Piaget's comments on coordination before. Coordination seems to be very similar to conceptual integration. I also found that you had told me that Michalski had a fast inferencing method so that must be important to you for some reason. So, to repeat myself for clarity. I am going to run a subjective experiment for a couple of weeks. The goal is to get you to solve a problem for me and I want to be able to note how I personally integrate subject related serendipity into my knowledge structures concerning the subject. It is unlikely that you would be able to solve a problem that I specified in advance so I am going to look for an unexpected serendipitous solution to some problem that I haven't yet completely identified. In order to get you to participate in this experiment I need to encourage you to talk about your project using terms that are relevant to both of us. Since I will be keeping notes I have started by reviewing and collecting some of the comments you made in this thread. I can then use this knowledge to get you to continue talking about things that interest you. I noted that you have not acknowledged that complexity is a problem so I will back off that particular problem and try to shift to integration (coordination) issues that seem challenging for an automated AGI program to use effectively. Now that I have explained this 'experiment' to you I will stop talking about it and get back to the subject. On the list of mental coordination methods, internal simulation methods and inferencing you did not specifically mention conditional branching so there is a chance that you (or Piaget) left that off the list. I would say that is a pretty important concept! On the other hand, running different methods to use in a comparison with perceived events seems to imply a conditional branching. Anyway, the next question I have for you concerns 'verification' and integration (coordination). Without strong verification, coordination is essentially going to tie weak inferences together. If you accept that this could be a problem then how would your program use the products of coordination reliably? Jim Bromer On Tue, Dec 4, 2012 at 11:45 PM, Piaget Modeler <[email protected]> wrote: "The central idea is that knowledge proceeds neither solely from the experience of objects nor from an innate programming performed in the subject, but from successive constructions, the result of constant development of new structures.” ~ Jean Piaget So I think we knit together these insights, piecemeal, until they recur and strengthen, and become more predictable and forceful in our minds. Then they integrate and form a larger structure, and eventually they become a subsystem, integrating with other subsystems, until they finally integrate with the totality. Or at least that's how I interpreted it in "The Development of Thought" by J.Piaget. Cheers. ~PM. Date: Tue, 4 Dec 2012 23:12:06 -0500 Subject: Re: [agi] Internal Representation From: [email protected] To: [email protected] Well, I would look at Ryszard Michalski's work on dynamically interlaced hierarchies if it was convenient for me to do so. Nothing about this is mentioned on his home page and the first reference I looked at did not seem like a breakthrough paper. I want to finish something that I was thinking about. We (or a machine) would be able to build strong knowledge if the knowledge that was gained could be used to reliably predict, explain or produce a specific outcome. But often, the outcomes are weak or unreliable indicators of much of value. So instead we are left with a lot of weakly related situation-action-reaction insights that are inexplicably conditional and variant. This is a lot like serendipitous learning. If I try to learn something, I probably won't be able to figure out what I wanted to figure out (unless it is something that other people had already figured out and it was within my field of knowledge). But I would probably learn something new serendipitously. Now can we patch a lot of weak unexpected insights together? Yes, but in order to build something reliable out of a lot of weak structural pieces they have to be integrated pretty thoroughly. The integration does not have to perfect but the matrix of these things have to be strong enough to serve as a foundation for greater insights. Jim Bromer On Tue, Dec 4, 2012 at 9:31 PM, Piaget Modeler <[email protected]> wrote: I would agree that you also need mult-strategy reasoning in addition to correlations. Look at Rysard Michalski's work on dynamically interlaced hierarchies. He has a fast and efficient mechanism for inference. He inspired me. Cheers, ~PM. Date: Tue, 4 Dec 2012 18:36:20 -0500 Subject: Re: [agi] Internal Representation From: [email protected] To: [email protected] I discovered something about logic that I never knew before. It is something that I have thought about for 40 years, but I never stopped to explore the application. Now, shouldn't this new insight give me greater understanding? Well, yeah, but it doesn't work that way. I have a new insight but I haven't got any use for it. So now I have to try to find some practical use for it. Well even though I don't have any use for it, I might pick up some street creds by telling other people about it right? Well no, not really. It is really a turn-the-crank kind of thing and the fact that I thought about it for so long without ever once examining its application is kind of embarrassing. So now, before I can talk about it I have to search for some way to use the idea effectively. If I found some utility for it then I could pick up some credit for it, but until then it is just going to make my work with logic more complicated. The insight was a turn-the-crank kind of insight so it represented the application of a familiar idea onto another familiar idea in a way that was very familiar to me. The only thing I did different was to actually see how it worked in a few examples. When I did that I realized that the effects were not exactly what I expected. However, logic is an artificial field which is well formed so that other logic-based ideas, like something from mathematics, can sometimes be easily integrated into it. In real world examples of ideative projection, the analysis of turn-the-crank imagination cannot easily be achieved just by using other (integrated or related) methods of internal ideative projection. And as I just explained, simple correlation methods are not an easy substitute for insightful methods. Jim Bromer AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription 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
