Neo-Modular Systems: Architectural Principles for Cognitive Robotics
Speakers: Lynn Andrea Stein
Topic(s): Artificial Intelligence
Abstract
Traditional computational systems are generally organized around principles of modular-functional decomposition. In contrast, cognitive robotics has produced a large number of apparently non-modular systems. These systems rely on opportunism: they reuse
solutions to simpler (or more primary) problems in ways which do not
fit traditional modular-functional decomposition. I will argue that
these opportunistic systems can be explained by and engineered with a
new set of modularity principles, in which functional decomposition
and modular decomposition may not be the same.
I will describe three concrete "neo-modular" principle:
o Imagination puts mental machinery for solving physically
interactive problems to use in purely cognitive domains.
o Shared Grounding aligns disparate representational systems --
builds representational transducers -- by exploiting their
abilities to process external ground truth.
o Incremental Adaptation builds complex behavioral solutions by
``growing'' simpler solutions to handle increasingly difficult
problems.
Each of these three enables the construction of a wide variety of
systems, and I will review several of these. I will then describe
ways in which the three approaches are intercompatible, and present a
speculative system which builds on all three principles to solve
problems of hand-eye coordination and reaching.
In systems built according to these principles, there is of necessity
no neat correspondence between the functional decomposition of the
system's behavior and any modular decomposition of its implementation.
It is my hypothesis that natural systems, and the most successful
approaches to cognitive robotics, will be in this way neo-modular.
About this Lecture
Duration: n/a minutes
Languages Available: English
Last Updated: 12-14-2007
Request this Lecture
To request this particular lecture, please complete this online form.
The request will be sent to ACM headquarters for review.
|
|
