Learning to Plan Near-Optimal Collision-Free Paths
A. Ho, Geoffrey Fox · 2005
A new approach to find a near-optimal collision-free path is presented. The path planner is an implementation of the adaptive error back-propagation algorithm which learns to plan “good”, if not optimal, collision-free paths from human-supervised training samples. Path planning is formulated as a classification problem in which class labels are uniquely mapped onto the set of maneuverable actions of a robot or vehicle. A multi-scale representational scheme maps physical problem domains onto an arbitrarily chosen fixed size input layer of an error back-propagation network. The mapping does not only reduce the size of the computation domain, but also ensures applicability of a trained network over a wide range of problem sizes. Parallel implementation of the neural network path planner on hypercubes or Transputers based on Parasoft EXPRESS is simple and efficient, Simulation results of binary terrain navigation indicate that the planner performs effectively in unknown environment in the test cases.