Assisted Research and Optimization of the Proper Neural Network Solving the Inverse Kinematics Problem

Abstract

Finding the better solution of the neural network design to solve the inverse kinematics problem with the minimum of the trajectory errors is very difficult, because there are many variable parameters and many redundant solutions. The presented paper show the assisted research of the influences of some more important parameters to the final end-effector trajectory errors of the proposed neural network model solving the inverse kinematics problem. We were been studied the number of neurons in each layers, the sensitive function for the first and second layer, the magnifier coefficient of the trajectory error, the variable step of the time delay and the position of this block, the different cases of target data and the case when the hidden target data were adjusted. All obtained results were been verified by applying the proper direct kinematics virtual LabVIEW instrumentation. Finally we were obtained one optimal Sigmoid Bipolar Hyperbolic Tangent Neural Network with Time Delay and Recurrent Links (SBHTNN(TDRL)) type, what can be used to solve the inverse kinematics problem with maximum 4% of trajectory errors.