LEIPZIG – Through an intricate series of algorithms, Martin
Riedmiller, PhD, explained how to teach a machine to learn. Riedmiller’s
research was conducted in an effort to address the growing need for intelligent
learning systems or software that adapts behavior based on experience, like
humans.
In a presentation to the 13th ISPO World Congress, here, he compared the
process to that of a baby learning in the early stages of infancy. However,
there is one large difference. Unlike a child, a machine can not combine
learned behaviors and will only follow one set of learned instructions.
 |
|
| Martin Riedmiller | |
“In machine learning … the main question is how can we relay
the stream of data that is coming in — for example from a camera or
artificial skin, to a stream of action data, which might be motor — in
order to achieve a certain goal?”
Riedmiller achieves this through a string of biological neurons weighted
with specific parameters adapted until the desired effect is reached. He used
three examples to explain this process: a cart balancing a rod, robot soccer
and slot car racing.
“The mechanics to build a mechanical arm, to build mechanical hands
are making big progress … so basically the input and output are already
there,” Riedmiller said. “What’s lacking is the algorithm for
the brain.”
The example of myoelectric control is one already where people are using
the mathematical techniques presented today in a practical way to control
prosthetics. [As an example] we had patients who were using a forearm with an
amputation and forearm without amputation to collect myoelectric signals from
both and present them to the neural network and it learns. Pretty soon, the
amputee is able to manipulate a wrist in two degrees of freedom and the hand
opening and closing. This is an example of the kind of things you are able to
do. I see a great opportunity now for these techniques to be increased in their
applicability to our control systems for prosthetic devices in general.
— T. Wally Williams, MA
Liberating
Technologies Inc.
