Page 119 - Socially Intelligent Agents Creating Relationships with Computers and Robots

102 Socially Intelligent Agents

behalf of the people or resources they represent, while also ensuring that the
selected team collectively possesses sufficient resources and capabilities. The
proxies could also monitor the progress of the participants and of the mission
as a whole, executing corrective actions when necessary.
Applications of agents within human organizations have fostered an increas-
ing interest in adjustable autonomy (AA), where agents dynamically adjust their
own level of autonomy, harnessing human skills and knowledge as appropriate,
without overly burdening the humans. When agents are embedded in large hu-
man organizations, they must also coordinate with each other and act jointly in
teams. The requirements of teamwork and coordination give rise to novel AA
challenges not addressed by previous research, which focuses on interactions
between only an individual agent and its human user [3, 4, 5]. In particular,
the AA coordination challenge arises during the transfer of decision-making
control. In a team setting, an agent cannot transfer control freely, because as
the agent waits for a human response, its teammates expect it to still fulfill its
responsibilities to the overall joint task. Thus, the AA coordination challenge
requires that an agent weigh possible team miscoordination while waiting for
a human response against possible erroneous actions as a result of uninformed
decisions.
We have conducted our research on AA using a real-world multi-agent sys-
tem, Electric Elves (E-Elves) [1], that we have used since June 1, 2000, at
USC/ISI. E-Elves assists a group of 10 users in their daily activities. To address
the AA coordination challenge, E-Elves agents use Markov decision processes
(MDPs) [8] to explicitly reason about team coordination via a novel three-
step approach. First, before transferring decision-making control, an agent
explicitly weighs the cost of waiting for user input and any potential team mis-
coordination against the cost of erroneous autonomous action. Second, agents
do not rigidly commit to transfer-of-control decisions (as is commonly done in
previous work), but instead reevaluate decisions as required. Third, an agent
can change coordination arrangements, postponing or reordering activities, to
“buy time” to lower decision cost/uncertainty. Overall, the agents look ahead
at possible sequences of coordination changes, selecting one that maximizes
team benefits.

2. Electric Elves
As a step towards agentization of large-scale human organizations, the Elec-
tric Elves effort at USC/ISI has had an agent team of 15 agents, including 10
proxies (for 10 people), running 24/7 since June 1, 2000, at USC/ISI [1]. The 5
other agents provide additional functionality for matching users’ interests and
capabilities and for extracting information from Web sites. Each agent proxy is
called Friday (from Robinson Crusoe’s servant Friday) and acts on behalf of its