Abstract: Complex Systems and AGI, By Richard Loosemore
|
|||||
|
|
|
||||
|
||||||||||||
|
Abstract posted for the May 20-21 AGI Workshop: Complex Systems and AGI An Extended Abstract By Richard Loosemore (former Director of Research at Starbridge Systems)  PPT: Complex Systems and AGIThe purpose of this talk is to make three claims. (1) To argue that the Complex Systems Critique of AI has often been seen as somewhat vague, whereas in fact it can be formulated in a way that makes it precise, powerful and prescriptive. (2) To propose a formalism for describing the structure of intelligent systems that exposes the weaknesses of AI that are targeted by the Complex Systems Critique. (3) To propose a software development environment designed specifically to allow intelligent systems to be built that are consistent with the suggested formalism. Complex Systems researchers argue that we should be careful about a trap that could be called the “global-to-local disconnect,” which works as follows (1) We start out with the intention of capturing the global properties of intelligence using systems of interacting, adaptive local elements. (2) Our interacting local elements turn out to constitute a complex system. (3) Because these systems are complex, there is not necessarily any obvious or analytic connection between local and global, so our naïve local implementation of the local mechanisms is not guaranteed to yield the global behavior we desire. (4) We (perhaps unconsciously) modify our proposed local mechanisms and our research goals in order to make it seem that our systems are working better than they are. In the first part of the talk, we look briefly at some of the ways that the global-to-local disconnect actually manifests itself in practice, and at some of the ways that we try to sidestep it. In the second part of the talk we propose a formalism for describing the structure of complete AI (AGI) systems. At the core of the framework is a “foreground” (working memory) in which “elements” (basic units of knowledge) can instantiate and interact with one another. In general, what these elements do is engage in a process of dynamic relaxation: they impose weak constraints on one another in an attempt to form connected structures of elements that are mutually consistent. External constraints act on the foreground from a sensory input area, while an analogous motor output area allows the foreground to have effects on the world. Most of the meat of the formalism is in the specific choice of local mechanisms used to implement the dynamic relaxation: we describe some of the candidate mechanisms, and examples of how various conventional AI systems map onto this framework. In the final part of the talk we suggest that a proper methodology for escaping the complex systems problem is to build a software development environment that allows AGI systems to be quickly built and systematically compared. The previously described conceptual framework is designed to be used as the language within which AGI systems would be built in this development environment. We outline both the structure of the proposed environment and the way it would be used as part of a concerted research paradigm. The main features of the proposed environment are its use of parallelism, the ease with which systems can be constructed, and the tools for empirical study of the behavior of candidate systems. Important Links: Main Workshop Website: http://www.agiri.org/workshop Directions/Hotel: http://www.agiri.org/directions.htm Workshop Schedule: http://www.agiri.org/schedule.htm Printable Version / Handout: http://www.agiri.org/workshop/AGIRI_Workshop_2006.pdf |
Posts in this topic
    |
|
AGIRI · Forums · Discussion Lists · Wiki |  |
AGIRI - Artificial General Intelligence Research Institute Copyright © 2001-2007 :: User Agreement :: Discussion Lists :: Contact |