Remarks on Control of Robot Manipulator Using Quaternion Neural Network
Kazuhiko Takahashi · 2018
High-dimensional neural networks, in which all the network parameters, states, signals and activation functions are expressed using hypercomplex numbers, have received increasing attention as solutions to real-world problems in many fields of science and engineering. Quaternion numbers constitute a class of the hypercomplex number system, and several successful applications based on quaternion neural networks have been demonstrated. In this study, the application of a quaternion neural network to control systems is investigated. An adaptive-type servo controller, in which a quaternion neural network output is used as the control input of a plant to ensure the plant output matches the desired output, is presented. The quaternion neural network has a multi-layer feedforward network topology with a split-type quaternion function as the activation function of neurons, and a tapped-delay-line method is used to compose the network input. To train the network parameters by using the gradient error minimisation, a feedback error learning scheme is introduced into the control system. Computational experiments for controlling a two-link robot manipulator by using the proposed quaternion neural network-based controller are conducted, and the simulation results obtained demonstrate the feasibility of using the proposed controller in practical control applications.