BLACK BOX CLOSED LOOP ROBOT MANIPULATOR SYSTEM IDENTIFICATION

The paper discusses experimental identification of one joint of a hand made, two degrees of freedom robot manipulator, including flexibilities, under feedback. A black box system model is identified from the input-output data. Both linear, OE (Output Error) and non-linear structure (multilayer perceptrons neural network) models are treated and applied. A Levenberg-Marquardt algorithm is implemented to generate our NNARX model. As regressors two past inputs and two past outputs are chosen. Furthermore network architecture is chosen with 5 hidden tanh units and one linear output unit. Fit criteria shows that the linear model has severe problems. Validation of the trained non-linear network looks quite satisfactory, and it is definitely better than the linear model. Experience has shown that regularization is helpful when pruning neural networks. A remarkable improvement in performance, when using long instead of short format for choosing neural network weights and Bias, is appreciated.