======== WDL Dis. #11 (Bcc: all WDL participants) ====== Date: Thu, 02 Mar 2000 15:49:02 +0000 From: Stan Franklin To: Dr John J Weng Subject: Re: WDL Dis. #9 John, Thanks for the reference. I, of course, know of the Robo-SOAR paper, but I haven't read it. > However, it seems that, among other fundamental problems, > a basic fundamental problem with Soar (as well as other systems) > is the use of SYMBOLIC REPRESENTATION. With a symbolic > representation, the representation must be assigned with meaning > (by human) at the time of programming. However, an algorithm > for automatic development must automatically generate > representation. For example, an infant who is born in US is > moved to China right after the birth will learn Chinese > successfully. If the developmental algorithm of the child > (in his gene) is pre-built with symbolic representation of > English, he will not be able to learn Japanese. Thus, it > seems that a system that can automatically develop MUST NOT > use symbolic representation. My agents sense text as humans sense light or sound waves. "Tues." sensed as a piece of text may initiate activity in the perception module. There it may activate a node labeled TUESDAY which may in turn activate a DAY-OF-THE-WEEK node that would activate several message type nodes that use this particular kind of information. Eventually TU may be written to the workspace (working memory) and to long term associative memory along with other information perceived from an incoming message. This isn't meant to be arbitrary symbolism, but rather grounded by perception. It's more akin to our recognizing an apple than to our recognizing the meaning of a word. To the agent TU isn't a symbol that represents something else. Rather, it's a partial internal state that contributes to causal effects on action selection. Of course, this grounding is based on knowledge contained in the codelets and the network of the perceptual module. In humans this kind of domain knowledge (words) must be learned and has to do with language. In these agents some initial knowledge is built in by the designers, just as some perceptual knowledge is built in to some animals through evolution (nestling birds ducking from a hawk's shadow). I believe that some amount of prior knowledge must be built in either by a designer or by evolution in order for any learning to occur. We can only learn as extensions from what we already know. If I'm wrong about this, I'd certainly be happy to be corrected via an example. > In a non-developmental system, it is the programmer who understands > the task(s) that the system will deal with. Think about a machine. > > In a developmental system, it is the system itself to understand > the task(s) that the system will deal with. Think about a human. I intend for my agents to be developmental systems, learning both concepts and behaviors in the course of their "lives". Our work along these lines is so far only in the planning stage, but we've written about it. Ramamurthy, Uma, Franklin, Stan, and Negatu, Aregahegn . "Learning Concepts in Software Agents." In Fifth International Conference of the Society for Adaptive Behavior-SAB98. Zurich, Switzerland, MIT Press, 1998. Bogner, Myles, Ramamurthy, Uma and Franklin, Stan. (In Press). "Consciousness" and conceptual learning in a socially situated agent. Dauthenhahn, Kerstin (Ed.). Human Cognition and Social Agent Technology. Amsterdam: John Benjamins Publishing Company. Stan -- Stan Franklin Math Sciences Dept Phone: (901) 678-3142 Univ of Memphis Fax: (901) 678-2480 Memphis, TN 38152 stan.franklin@memphis.edu USA www.msci.memphis.edu/~franklin