- Title:
- Learning to Fly by Controlling Dynamic Instabilities
- Presenter:
- David Stirling
- Date:
- August 09, 2002
- Abstract:
- This talk is about a new approach for learning and building
computational models of human skills applied in complex control
situations. Such skills have usually been internalised as automatic and
possibly sub-cognitive responses, such as driving a car. In the recent
past a degree of success in modelling these has been partly addressed
with behavioural cloning. Skills obtained by this technique, often
exhibit lack generality and robustness in new or only slightly
unfamiliar situations. This is mitigated in the new approach by
segmenting and compressing a universal set of reaction plans with
symbolic induction methods. This approach is termed, Compressed
Heuristic Universal Reaction Planners, or CHURPs. CHURPs provides a
substantially improved robustness and control performance, this arises
from synergistic interactions and collaborations between the different
CHURP units, including, surrogate control and goal sharing. The results
observed also mark a similarity to the behaviours observed in the field
of ethology ¾ the study of human behaviour in terms of patterns of
muscle activity.
Back to:
|
