I like Aaron's solution. In fact, all the blue boxes in the PAM-P2 architecture diagram http://piagetmodeler.tumblr.com represent areas in the knowledge-base and all the other boxes represent specialized agent programsserving as "functions" over the KB. I think that's a trending pattern these days. Agents acting upon a knowledge base or data base. My $0.02 ~PM
Date: Tue, 11 Dec 2012 21:17:04 -0600 Subject: Re: [agi] Internal Representation From: [email protected] To: [email protected] But that doesn't explain *what* would use it. People use Search engines but when we go to imagining how an AGI program would use a knowledge index graph we have to conclude that either the algor-unculus is separate from the knowledge graph or that its actions can be derived from knowledge graph. While the subprogram (the agent that I am imagining would 'use' the knowledge graph) would be programmed with some default values I see this as also being able to learn. It is this ability, the ability to truly learn something, and use that knowledge as the basis for judgement, that would make the program act like it was capable of understanding. So the knowledge index graph would include insights about using the knowledge index graph when making certain kinds of searches so that the algor-unculus would be able to recognize that some information could be translated into (algorithmic) actions that it could take in trying to 'understand' a problem. I would lean more towards the idea of having multiple cooperative agents or inference rules, each storing its state not internally, but in the shared "knowledge index graph" (a.k.a. semantic net) and being triggered by relevant changes made by the others. A set of such agents or rules could be evolved or otherwise learned automatically by comparing the nodes/vertices or connections/edges they create against those produced from direct observation (or other agents/rules) to determine their efficacy at modeling the real world consistently. The meta-algorithm which is used to produce, evaluate, modify, and cull these agents/rules would be where the intelligence comes from (learning), but the agents/rules would collectively make up the actual intelligence of the system (competence). On Tue, Dec 11, 2012 at 2:11 PM, Jim Bromer <[email protected]> wrote: I was just asking Google and Bing questions and I was surprised at how well they did. No, they were not able to answer more complicated questions or refine on their searches if I did not know how to express my simple questions with more refined key words but compared to the Search engines of 10 years ago they are amazing. In some of the cases my question was echoed as a key phrase in the website that was indexed, but that was not always the case. So this convinces me that contemporary search engine technology is not just narrow AI although it is not general AI either. http://www.google.com/insidesearch/features/search/knowledge.html A knowledge graph is pretty much what I was getting at when I mentioned index branching. I realize now that the word "graph" was a better word. I would use a knowledge graph where the nodes can contain distributed 'conceptual information' related to other nodes or index information that showed how some group of 'concepts' were related. I only used the term branching or tree to emphasize that by shaping how this graph of interrelated concepts is used might isolate certain information if the search conditions seemed to merit it. This would avoid the combinatorial explosion for some cases. So each node of the knowledge graph might only contain index information, but they might contain more than one kind of index. Or perhaps there might be some system that governed the use of the graph so that different kinds of searches could use different kinds of methods to search the data. So I would use a knowledge index graph which could be governed by different methods of using it to search for information. But that doesn't explain *what* would use it. People use Search engines but when we go to imagining how an AGI program would use a knowledge index graph we have to conclude that either the algor-unculus is separate from the knowledge graph or that its actions can be derived from knowledge graph. While the subprogram (the agent that I am imagining would 'use' the knowledge graph) would be programmed with some default values I see this as also being able to learn. It is this ability, the ability to truly learn something, and use that knowledge as the basis for judgement, that would make the program act like it was capable of understanding. So the knowledge index graph would include insights about using the knowledge index graph when making certain kinds of searches so that the algor-unculus would be able to recognize that some information could be translated into (algorithmic) actions that it could take in trying to 'understand' a problem. Although I believe that the complexity problem is too great a problem for me to solve, I do believe that I could demonstrate what I am talking about with simplistic idea-world that could stand as evidence that this is a feasible model (given more computing power). So this is a statement of an experimental test that can be evaluated. Most of us would be able to recognize whether or not I (or someone) was able to use these ideas with a simple data world or not. Jim Bromer 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
