Project has two overriding goals: 1) to develop distributed
agent-based architectures that are composed of negotiating
and learning agents, and 2) to apply these architectures to
everyday activity management and information problems. To
meet these goals, we have developed a number of software agent
systems, including agents for managing personal calendars,
web agents that provide guided tours of websites, and a visitor
hosting agent system that manages the process of connecting
faculty members with campus visitors who share similar interests.
is committed to developing machine learning methods that will
endow personal software agents with the ability to customize
automatically to the needs of their users.
The Calendar Apprentice (CAP) is an interactive assistant
that acquires knowledge through routine use by observing users'
actions. CAP provides an editing and email interface to an
online calendar. It learns users' scheduling preferences through
routine use, enabling it to give customized scheduling advice
to each user. Each night, CAP automatically runs a learning
process--based on a method of decision tree induction--in
order to refine the set of rules it uses to give scheduling
advice. One of the advantages of this approach to machine
learning is that the rules in CAP are typically understandable
to users, which may encourage user editing and evaluation.
WebWatcher, a learning apprentice like CAP, is a tour guide
agent for the Web that accompanies users and offers suggestions
about where to go next. Given that users often feel disoriented
when encountering websites for the first time, the WebWatcher
acts as a guide, offering suggestions based on its knowledge
of each user's interests, the location and relevance of various
texts available, and the ways in which previous users have
interacted with the website. As WebWatcher accompanies the
user, it offers suggested links by adding eyeball icons to
its selections. The WebWatcher agent learns to suggest appropriate
links for users by analyzing the training examples of previous
tours, in which user-selected links are annotated with the
keywords of users who chose them. A second learning strategy
involves reinforcement, in which the agent augments a given
link using words encountered in pages downstream of it. A
third approach to learning combines these two approaches,
and has been shown to be more effective than other learning
The Visitor-Hoster system is designed to help a human secretary
organize a visit in an academic environment. According to
the demands of this task, the visitor's schedule must be arranged
to accommodate the schedules of faculty members who share
interests with the visitor. We have developed a layered architecture
to meet the demands of hosting a campus visitor. This architecture
deploys task-specific software agents that help users
perform tasks by communicating with each other and/or querying
and exchanging information with information-specific
agents. Task agents include a Personnel Finder agent that
locates information about the visitor, a Scheduling Task agent
that manages the visitor's calendar of meetings with faculty,
and various Calendar Apprentice (CAP) agents that manage and
interact with faculty calendars. Information agents engage
databases and other sources of information, as in the Interest
Agent that locates faculty members whose interests match the
visitor's areas of interests.
is a successful
application of Retsina multi-agent technology that integrates
three sources of information to make a prediction of satellite
coordinates of the region of observation;
forecast websites for the region of observation;
of visible satellites over the area at the specified time.
for a demonstration.
information, visit the Pleiades
Project Home Page.