Managing Semantic Compensation in a Multi-Agent System
Amy Unruh
|
Computer Science & Software Engineering
|
University of Melbourne
|
Friday, 25th June 2004
9-30am - 11am
RMIT University Function Room (near Kaleide Theatre)
Building 8, level 2, RMIT,
Swanston Street, Melbourne.
Abstract:
This talk will describe an approach to improving the robustness of an
agent system by augmenting its failure-handling capabilities. The
approach is based on the concept of "semantic compensation": undoing
or "cleaning up" failed or cancelled tasks can help agents behave more
robustly and predictably at both an individual and system level.
However, in complex and dynamic domains it is difficult to define
useful specific compensations ahead of time. This talk will
present an
approach to defining semantic compensations abstractly, then
implementing them in a situation-specific manner at time of failure.
The talk will then describe a methodology for decoupling
failure-handling from "normative" agent logic so that the semantic
compensation knowledge can be applied in a predictable and consistent
way -- both with respect to individual agent reaction to failure, and
with respect to failure-related interactions between agents --
without requiring the agent application designer to implement the
details of the failure-handling model.
Biosketch:
Amy Unruh is an ARC Research Fellow at the University of Melbourne
Department of Computer Science and Software Engineering.
Recently,
she has been an adjunct faculty member at the University of Texas, has
worked for various startups, and was a member of the InfoSleuth
Distributed Agents Program at MCC in Austin. She received her
Ph.D. in CS/AI from Stanford University.
Papers:
A. Unruh., J. Bailey, and
K. Ramamohanarao. "A Framework for Goal-Based Semantic Compensation in
Agent Systems." To appear in 1st International Workshop on Safety and
Security in Multi-Agent Systems, AAMAS '04, 2004 (local pdf).
A.
Unruh., J. Bailey, and K. Ramamohanarao. Managing Semantic
Compensation in a Multi-Agent System (pdf, submitted for
publication).