Remarks on Feedforward-Feedback Controller Using a Trained Quaternion Neural Network Based on Generalised HR Calculus and Its Application to Controlling a Robot Manipulator

In this study, a gradient-descent method extended to quaternion numbers, in which the generalised Hamiltonian-Real calculus is used to calculate derivatives of a real function with respect to quaternion variables, is explored. It was used to derive a training algorithm of a quaternion neural network that functions as an adaptive-type controller in control systems applications. A feedforward-feedback controller is designed based on the quaternion neural network. Furthermore, computational experiments on trajectory tracking control of a three-link robot manipulator are conducted to verify the learning capability and the characteristics of the quaternion neural network-based controller. The experimental results showed the quaternion neural network's feasibility and capability for a control systems application.