FRC Seminar Announcement
			========================

Date:   Friday, Dec 13
Time:   1:30pm - 2:30 pm
Place:  Frc 200 (call X7089 to enter building from outside)

Etiquette: Gather 15 minutes early if you want to interract with other
	   attendees.
	   Do not arrive late.

Speaker: Bruce Digney, 
	 Defense Research Establishment, Canada

Title: Learning Reactive Hierarchical Control Structures

Abstract:

  The use of externally imposed hierarchical structures to reduce the
  complexity of learning control is common. However, it is
  acknowledged that learning the hierarchical structure itself is an
  important step towards more general and less bounded learning in
  which the robot  learns as many things as warranted as opposed to a
  single specified skill. Presented in this discussion is a reinforcement
  learning method that generates a hierarchical control structure. 
  Effectively, the learned hierarchy decomposes what would otherwise
  be a monolithic evaluation function into many smaller evaluation
  functions. These can be thought of as skills that can later be invoked
  as the robot learns new tasks and allows the reuse of previously
  learned sensory-motor information. When the emergent hierarchical
  structure is combined with learned bottom-up reactive responses, a
  reactive hierarchical control system results. The effect of this is
  twofold. Firstly, when learning a new but related task, previously
  learned skills that have been recognized as valuable are distinguished
  from primitive actions and many other irrelevant skills.  That is, when
  presented with a new task to learn, meaningful skills will "jump out of
  the clutter" for the robot's consideration. The robot can then view its
  world in terms of abstracted skills rather than only as primitive actions. 
  Secondly, when learning a new task and a situation appears that has
  previously proven to be hazardous or opportunistic, the appropriate
  skills will be reactively invoked to benefit the robot without the need
  for it to relearn an appropriate response. This avoids any damage or
  missed opportunities that might occur during relearning. In general,
  the continual carrying forward of learned skills and reactions will
  allow robots to accomplish increasingly more complex tasks. It also
  lends itself to the preparation of robots for their intended tasks using
  regimented training to allow robots to get a head start at learning
  complex tasks without impairing their autonomous learning
  capabilities. This discussion will present examples of learning
  hierarchical control structures, relevant feature emergence, a
  comparison with non-hierarchical learning systems and the use of
  continuously valued states, primitive actions and skill invocation
  strengths.