|
October
22 , 2002
 In
"Alpha Wolf," Alan Alda
checks in with Bruce Blumberg
during a return visit to The Media Lab at MIT. Blumberg has
long been inspired by dogs in his efforts to build autonomous
animated creatures that can learn. As Alan has witnessed firsthand,
Blumberg's virtual canines can be trained to respond to voice
commands just like a real animal. But why has Blumberg repeatedly
chosen dogs as his model? A dog-lover and owner of a silky
terrier named Sydney, Blumberg believes that dogs possess
the ideal qualities we humans look for in a companion, virtual
or real. So perhaps "man's best friend" has something to teach
us about how to make computers more sociable.
- - - - - - - - - - - -
Meeting
Expectations
People
expect a level of common sense from any animate system - that
it will "get the good" and "avoid the bad," given its desires,
repertoire of actions, and beliefs about how the world works.
Part of the common sense expected from an intelligent system
is that it will learn from experience. Indeed, the ability
to learn from experience is one measure of what people often
label as intelligence, i.e., more intelligent creatures are
better able to learn than less intelligent ones. When a character
doesn't learn from experience, we are left wondering "is it
stupid, or is it simply broken?"
 |
 |
|
Blumberg's
goal is to make virtual dogs like Duncan as trainable
as a real dog.
|
|
The
goal of the Synthetic Characters Group at the Media Lab of
MIT is to understand how to build autonomous animated characters
that possess the everyday common sense, ability to learn,
and apparent sense of empathy that one finds in animals such
as dogs. That is, we take our fundamental inspiration from
animal learning and training. Our belief is that by paying
close attention to how animals learn and to successful techniques
by which they are trained, we can not only improve on existing
models for machine learning, but also develop robust techniques
for real-time learning in autonomous animated characters.
-
- - - - - - - - - - -
3
pages: | 1 | 2 | 3
|
|
