Fractals
show machine intentions
By Eric Smalley, Technology Research News
There has been much research and musing about how autonomous
machines like robots and intelligent software agents should interact
with people. Much of the work focuses on giving machines a degree
of social intelligence that will allow people to understand and communicate
with them on human terms.
A sense of internal states is integral to human communications:
it's useful to have a sense of when a human is annoyed. In contrast,
it's often impossible to determine whether a robot is processing data,
awaiting instruction or in need of repair.
Researchers from Switzerland and South Africa have designed
a visual interface that would give autonomous machines the equivalent
of body language.
The interface represents a machine's internal state
in a way that makes it possible for observers to interpret the machine's
behavior. "Our idea of communication has a strong focus on learning
and interpretation -- trying to create relationships between the internal
machine variables and the macroscopic behavior," said Jan-Jan
van der Vyver, a researcher at the University of Zürich and the
Swiss Federal Institute of Technology.
The researchers' autonomous machine interface consists
of a clustering algorithm that groups the machine's many internal
states into a manageable number of representations, and a fractal
generator.
Clustering algorithms organize data like that contained
in genes into groups with similar traits, and analyze raw data without
any sense of the data's meaning or assumptions about how it should
be structured.
In the researchers' scheme, snapshots of a machine's
sensory input, computational processing and output are clustered and
the clusters are displayed as fractal images. The fractal generator
produces a fractal pattern in the center of the display and patterns
move outward in concentric rings, giving observers a sense of change
over time.
Fractal generators produce a large variety patterns
that people are quick to distinguish. A set of snapshots corresponding
to a high degree of sensory stimulation could be clustered into a
representation of the machine that people learn to associate with
the machine observing a change in its environment, for example.
In coming up with a way to convey the data, the researchers
were careful to avoid any anthropomorphic representations that human
observers might associate with particular behaviors or intentions,
according to van der Vyver. Those associations are not likely to correspond
to the machine's behavior, he said.
The fractal display served as the interface to a neural
network that controlled the input and output devices of a smart room
at the Swiss national exposition Expo.02 from May to October 2002.
Exposition goers were able to interact with the room through the room's
cameras, microphones, pressure sensors, light projectors and speakers.
Observers were able to correlate the room's behavior
with the fractal display, said van der Vyver. "What we found
surprising was that the general public so quickly gravitated toward
our chosen implementation of the communication interface, and so quickly
learned to interpret it," he said.
The smart room, dubbed Ada -- The Intelligent Space,
was not a fully autonomous system, but demonstrated the viability
of the fractal display, said van der Vyver. Truly autonomous systems
are likely to emerge in the future, in part due to self-developing
technologies like genetic algorithms that evolve optimized designs,
he said.
Given the prospect of self-evolving machines, the researchers
argue for a broad definition of autonomous systems as systems developing
according to their own dynamics through interaction with their environment.
The ultimate in autonomous machines is a system that develops intelligent
behavior simply as a result of participating in a society, van der
Vyver said.
It's not clear that the researchers' approach is necessary,
said Jeffrey Nickerson, an associate professor of computer science
at Stevens Institute of Technology. Autonomous machines could be programmed
to explicitly represent their intentions, he said. "If understanding
intentions is hard, then why not force the machine to provide indications
of intentions, or at least a trace of reasoning?"
Initial practical applications of the researchers' work
are about five years away, said van der Vyver. "As the development
and deployment of more autonomous machines takes place, this research
comes into play," he said. However, self-evolving, self-repairing
machines are a long way off, he said.
Van der Vyver's research colleagues were Markus Christen
and Thomas Ott of the University of Zürich and the Swiss Federal
Institute of Technology, Norbert Stoop of the Swiss Federal Institute
of Technology, Willi-Hans Steeb of the Rand Afrikaans University,
the International School for Scientific Computing in South Africa
and the University of Applied Sciences of Northwestern Switzerland,
and Ruedi Stoop of the Swiss Federal Institute of Technology and the
University of Applied Sciences of Northwestern Switzerland. The work
appeared in the March 31, 2004 issue of Robotics and Autonomous Systems.
The research was funded by the researchers' institutions.
Timeline: 5 years
Funding: University
TRN Categories: Robotics; Human-Computer Interaction
Story Type: News
Related Elements: Technical paper, "Towards Genuine Machine Autonomy,"
Robotics and Autonomous Systems, March 31, 2004
