4 October 1993, 3:00, WeH 4601


 Learning the Backtracking Policy for Potential-Guided Path Planning
			    Goang-Tay Hsu


A global potential function on grid representation is easy to compute
when the dimensionality of the space is small.  For higher
dimensionality, control points are introduced and the potentials are
computed in work space.  For arbitrarily shaped objects, local minimal
becomes the major source of inefficiency for path planning on grid
potentials in work space.  Speed-up can be achieved by avoiding going
into local minimal or recovering intelligently from them.  In this
talk, a method to learn to backtrack intelligently from local minima
is proposed.  Past experiences on backtracking are learned.  They are
recalled to suggest best back tracing steps when a local minimum is
encountered.  The potential well bottomed on the local minimum is then
marked to prevent further exploration.

GT