Georg Martius
Max Planck Institute, Leipzig
Information driven self-organization of robotic behavior
Wednesday 06th of August 2014 at 12:00pm
560 Evans
Autonomy is a puzzling phenomenon in nature and a major challenge in
the world of artifacts. A key feature of autonomy in both natural and
artificial systems is seen in the ability for independent
exploration. In animals and humans, the ability to modify its own
pattern of activity is not only an indispensable trait for adaptation
and survival in new situations, it also provides a learning system
with novel information for improving its cognitive capabilities, and
it is essential for development. Efficient exploration in
high-dimensional spaces is a major challenge in building learning
systems. We propose to implement the exploration as a deterministic
law derived from maximizing an information quantity. More
specifically we use the predictive information of the sensor process
(of a robot) to obtain an update rule (exploration dynamics) of the
controller parameters. To be adequate in robotics application the
non-stationary nature of the underlying time-series have to be taken
into account, which we do by proposing the time-local predictive
information (TiPI). Importantly the exploration dynamics is derived
analytically and by this we link information theory and dynamical
systems. Without a random component the change in the parameters is
deterministically given as a function of the states in a certain time
window. For an embodied system this means in particular that
constraints, responses and current knowledge of the dynamical
interaction with the environment can directly be used to advance
further exploration. Randomness is replaced with spontaneity which we
demonstrate to restrict the search space automatically to the
physically relevant dimensions. Its effectiveness will be
presented with various experiments on high-dimensional robotic system
and we argue that this is a promising way to avoid the curse of
dimensionality. This talk describes joint work with Ralf Der and Nihat Ay.(video)
Join Email List
You can subscribe to our weekly seminar email list by sending an email to
majordomo@lists.berkeley.edu that contains the words
subscribe redwood in the body of the message.
(Note: The subject line can be arbitrary and will be ignored)