Neural network structures for identification of nonlinear dynamic robotic manipulator

Abstract

Nonlinear dynamics robot manipulator control is much more complicated when compared with the objects having linear dynamics properties or non-linear static properties. Therefore, selection of the controller for the object having nonlinear dynamic properties seems to be an issue of paramount importance. In this paper, the authors have built a controller based on artificial neural networks. There are many ways to make choices among different network selection strategies; and in order to find the most optimum neural network, it requires a large number of tests being conducted on the objects, calculation of the number of layers in those objects as well as prediction of neural number in each layer. An ideal neural network should have lowest network training time as well as smallest error. Although, more number of neurons per layer means a reduction in error, this would create a difficulty in computation of neuron's parameters of the layer as well as cause an increment in network training time. Finding an optimum number of neurons per layer is therefore an important criterion. In this paper the authors have presented their experiences about network options and methods for balancing problems.