Identification of robot manipulators using neural networks and genetic programming
Kazuo Kiguchi, Hyeon-Ho Jang, Toshio Fukuda · 2003
An evolving neural networks (NNs) based identification method is proposed using genetic programming (GP). Advantages of both NNs and GP are combined in the proposed method. Identification of unknown/uncertain robot manipulators is realized by using the adaptation ability of NNs, and the architecture of the NNs is evolved by using the GP technique. Consequently, evolution of the NN architecture and adaptation of its weights are carried out in the proposed method. In the proposed GP, the architecture of each individual in the population is the same as a NN. The adaptation process of each NN (each individual in the population) to the unknown/uncertain robot manipulator is carried out using the back-propagation learning algorithm during the fitness evaluation process of each NN in GP. Therefore, the NN, which shows better adaptation to the unknown/uncertain robot manipulator, results in better fitness in the proposed method. In order to avoid frequent disruption of the important subtree of the NN caused by crossover operators, the worst subtree of one selected NN is replaced with the best subtree of the other NN. The back-propagated errors during the adaptation process are used for evaluation of the subtrees of each NN. This strategy makes the evolution of NN architecture more efficient than traditional GP. The effectiveness of the proposed identification method has been evaluated with a 2DOF planar robot manipulator.