A basis function approach to scheduled locally weighted regression for on-line modeling of nonlinear dynamical systems

This paper proposes a new scheme of on-line identification for feedforward (FF) learning control of an unknown nonlinear multi-input multi-output (MIMO) plant free of zero dynamics. This is achieved by constructing a FF controller consisting of a bank of linear approximation models for various operating points, which are discretized and called scheduler. Conventional schemes used piecewise constant/linear interpolation techniques to address the discretization. However, the accuracy of response shaping was insufficient. To improve the performance, we propose to take a basis function approach to tune the parameter of the FF controller. To verify the effectiveness of the proposed scheme, numerical simulation is carried out using the motion equation of a two-link manipulator.