Personalized Agents: Call for Participation

Although the term "agent" has come to mean many things, it perhaps has the most traction when identified with an anthropomorphized and autonomous program that acts as a personal assistant to a specific user (or set of users). In this model, the agent usually "lives" in a virtual world, may have access to data about its user, and is empowered to act on its user's behalf in a variety of computer-based tasks, including appointment scheduling, vetting messages, engaging in negotiation with other users, discovering items of interest, and even initiating contact with other users and agents.

This personalized agent has several qualities, including: (1) The agent is almost always "on," working on the user's behalf even when the user is not present; (2) The agent must continually adapt over a longtime horizon to the user's changing needs; and (3) The agent is mostly focused on modeling the behavior and preferences of a specific individual or small group of individuals, rather than discovering large trends over aggregrate data

We believe that a key component to a successful personalized agent is learning from the user, both supervised and unsupervised. This symposium will focus on the unique issues that arise when developing adaptive learning methods for such agents. In particular, we are interested in several topics:

• Unusual or non-standard domains where learning agents have an impact
• Balancing the possibly competing individual interests of many different users
• Learning methods that directly address the sparse data problem
• Learning methods that handle distinct changes in a user's preferences
• Natural ways of teaching agents specific behaviors
• Benefits of learning agents over agents with fixed behavior
• Evaluation of the effectiveness of individual learning strategies (e.g., case-based, explanation-based, inductive, reinforcement learning), or multistrategy combinations.
• Characterization of learning and adaptation methods in terms of modeling power, communication abilities, knowledge requirements and processing abilities of individual agents
• Agents learning by observing other agents
• Investigation of teacher-student relationships between agents and users
• User modeling

Submission Requirements

Those interested in participating should e-mail Charles Isbell at aaai2002-fs@ai.mit.edu by the submission deadline. The email should include keywords, and complete contact information for the authors, as well as the URL of either a brief statement of interest (1 page), or a complete paper (8 pages). The statement of interest or paper should be in PostScript, PDF or HTML.

Organizers

Charles Isbell, Michael Littman, and Peter Stone

Complete details regarding the AAAI Fall Symposium are available at: http://www.aaai.org/Symposia/Fall/2002/fss-02.html.