A Method of Error Driven Neuron Placement for Trajectory Learning

Learning smooth sequences of multi dimensional values trajectories can be performed by a neural network approximating a function which maps the one dimensional input space time onto anm dimensional output space workspace of the trajectory The underlying structure of the proposed network and the basic learning scheme for trajectory storage has already been published in Go erke et al This paper describes the policy of dividing the input space T into parcels each con ned to one single P Neuron using two error driven heuristics for parcel placement and adaptation of parcel boundaries Parcel Merging Splitting and Sliding Parcels