Page 120 - Socially Intelligent Agents Creating Relationships with Computers and Robots

Electric Elves 103

user in the agent team. If a user is delayed to a meeting, Friday can reschedule
the meeting, informing other Fridays, who in turn inform their human users.
If there is a research presentation slot open, Friday may respond to the invi-
tation to present on behalf of its user. Friday can also order its user’s meals
and track the user’s location. Friday communicates with users using wireless
devices, such as Palm Pilots and WAP-enabled mobile phones, and via user
workstations. We have used Friday’s location reasoning to construct a People
Locator that publishes the whereabouts of members of our research group on
a Web page. This automatically updated information provides a cheap means
for increasing social awareness (similar to previous work in the field [12]).
AA is of critical importance in Friday agents. Clearly, the more autonomous
Friday is, the more time it saves its user. However, Friday has the potential to
make costly mistakes when acting autonomously (e.g., volunteering an unwill-
ing user for a presentation). Thus, each Friday must make intelligent decisions
about when to consult its user and when to act autonomously. Furthermore,
Friday faces significant, unavoidable uncertainty (e.g., if a user is not at the
meeting location at meeting time, does s/he plan to attend?).
In addition to uncertainty and cost, the E-Elves domain raises the AA coor-
dination challenge. Suppose that, when faced with uncertainty, a Friday agent
consults its user (e.g., to check whether the user plans to attend a meeting), but
the user, caught in traffic, fails to respond. While waiting for a response, Fri-
day may miscoordinate with its teammates (other Friday agents), since it fails
to inform them whether the user will attend the meeting. This, in turn means
that other meeting attendees (humans) waste their time waiting. Conversely,
if, to maintain coordination, Friday tells the other Fridays that its user will not
attend the meeting, but the user does indeed plan to attend, the human team suf-
fers a potentially serious cost from receiving this incorrect information. Friday
must instead make a decision that makes the best tradeoff possible between the
possible costs of inaction and the possible costs of incorrect action.

3. Decision-Tree Approach to AA

Our first attempt at AA in E-Elves was inspired by CAP [7], an agent system
for helping a user schedule meetings. Like CAP, Friday learned user prefer-
ences using C4.5 decision-tree learning [9]. Although initial tests were promis-
ing [11], when we deployed the resulting system 24/7, it led to some dramatic
failures, including:

1 Tambe’s Friday incorrectly, autonomously cancelled a meeting with the division director.
C4.5 over-generalized from training examples.
2 Pynadath’s Friday incorrectly cancelled a meeting. A time-out forced the choice of an
(incorrect) autonomous action when Pynadath did not respond.
3 A Friday delayed a meeting almost 50 times, each time by 5 minutes, ignoring the
nuisance to the rest of the meeting participants.