Estimating movements of a robotic manipulator by hybrid constructive neural networks

Abstract We have investigated four constructive neural network algorithms for evaluating the movements of a robotic manipulator. During the course of the study, the algorithms were improved by merging regularization approach and architectural flexibility into the fully cascaded purely constructive algorithms. The numerical simulations were carried out by using two regression problems concerning estimation of kinematics and dynamics of a robot. The results show that the modified hybrid versions of the algorithms produce networks that reach better generalization performance with smaller number of hidden layers, parameters, and floating point operations than those created by the standard versions of the same algorithms.