A 3D transformation of a rigid link system using back propagation

Summary form only given. In this study, a rigid link system with three degrees of freedom was used to assess the ability of a neural net to generalize a transformation between the xyz end-point coordinates of the rigid body and the corresponding joint angles theta /sub 1/, theta /sub 2/ and theta /sub 3/. For this application, distortion can be quantified as a radial displacement between the expected and observed output patterns. A series of training point densities, within the input pattern domain, was used to characterize the interpolative properties of the net. A minimum averaged error term of 3.5 parts per thousand was calculated using a net with 32 middle layer units.<>