Learning Concepts in Software Agents

Document Type: 
Artificial intelligence
Animal Cognition
Date of Issue: 
R. Pfeifer, B. Blumberg, J.-A. Meyer, and S. W. Wilson
Journal/Publication Title: 
Proceedings of The Fifth International Conference on Simulation of Adaptive Behavior
MIT Press
Place of Publication: 
Mass, USA
Official URL: 
This concept-paper explores issues related to learning new concepts in software agents which inhabit dynamic domains. We argue that agents learn based on what they already know and agents solve new problems which they encounter by making analogies to previously solved problems of similar type. We explore these issues within the scope of the Cognitive Agent Architecture and Theory (CAAT) [Franklin, 1997]. The architecture of a multi-agent system, CMattie which is based on the CAAT strategy and inhabits an email-based dynamic domain, is described. CMattie gathers information from humans, composes announcements of next week's seminars, maintains a mailing list, mails the weekly seminar announcement to members of that mailing list, and learns new variations to seminars and other such events which have to be announced to the members of the mailing list as her domain changes with new types of seminar-like events. Learning mechanisms being implemented in this system which enable CMattie to adapt to her changing domain are described. Through her learning, CMattie acquires new domain-specific concepts, thus adapting to her dynamic domain.