Robocup Symposium Talk 1
Tuesday, July 19, 15:40-16:30 (International Conference Hall)
An Adaptive Architecture for Physical Agents
Pat Langley
Computational Learning Laboratory
Center for the Study of Language and Information
Stanford University, USA
Abstract
In this talk I describe Icarus, an architecture for physical
agents that unifies ideas from a number of traditions. The framework supports
the reactive execution of routine skills, but it combines this ability with
conceptual inference to ensure informed behavior, and it associates skills with
goals to ensure relevant action. The architecture defaults to execution whenever
it has an applicable skill that will achieve its current goal, but it falls
back on means-ends problem solving when it encounters an impasse. Moreover,
successful problem solving leads to the creation of new executable skills that
let the system achieve its goals more directly in the future. Thus, Icarus incorporates
insights from research on human problem solving, reactive control, logical inference,
and analytical learning. The architecture differs from predecessors in its focus
on physical agents, its commitment to hierarchical organizations of knowledge,
and its approach to the cumulative acquisition of these structures. The framework
is broadly consistent with results from cognitive pychology, but also providesa
realistic platform for developing adaptive intelligent agents. I illustrate
Icarus' capabilities on tasks from an in-city driving environment and report
experiments on this and other domains. This talk describes joint work with Nima
Asgharbeygi, Dongkyu Choi, Negin Nejati, Seth Rogers, Stephanie Sage, and Daniel
Shapiro.