Preorganized neural networks: Error back-propagation learning of manipulator dynamics

In estimation of some mapping by neural networks, a part of nonlinear functions included in the mapping to be learned is often known beforehand. For example, the equation of motion of the manipulator includes particular nonlinear functions such as sinusoidal functions and multiplication. The present article discusses the method used to embed known nonlinear functions into the error backpropagation neural network to utilize the knowledge in terms of the mapping to be learned. The network proposed is able to learn the known part by using the preorganized layer and the unknown part by using the hidden layer separately. Then the network is applied to the learning of the inverse dynamics of the direct-drive manipulator. When the preorganized layer is prepared corresponding to the equation of motion, the experimental results show that the network can improve the learning speed and the generalization ability and also can acquire the internal representation.