Enhanced PID: Adaptive Feedforward RBF Neural Network Control of Robot manipulators with an Optimal Distribution of Hidden Nodes
This paper focus on three inherent demerits of adaptive feedback RBFNN control with lattice distribution of hidden nodes: 1) The approximation area of adaptive RBFNN is difficult to be obtained in priori; 2) Only partial persistence of excitation (PE) can be guaranteed; 3) The number of hidden nodes is the exponential growth with the increase of the dimension of the input vectors and the polynomial growth with the increase of the number of the hidden nodes in each channel which is huge
... y for the high dimension of inputs of the RBFNN. Adaptive feedforward RBFNN control with lattice distribution of hidden node can improve solve the demerits 1) but just improve demerits 2) and 3) slightly. This paper proposes an adaptive feedforward RBFNN control strategy with an optimal distribution of hidden nodes. It solves the demerits 2) and 3) that the standard PE can be guaranteed and the number of hidden nodes is linear increase with the complexity of the desired state trajectory rather than the exponential growth with the increase of the dimension of the input vectors. In addition, we articulate that PID is the special case of adaptive feedforward RBFNN control for the set points tracking problem and we named the controller is enhanced PID. It is very easy tuning our algorithm which just more complex than PID slightly and the tuning experience of PID can be easily transferred to our scheme. In the case of the controller implemented by digital equipment, the control performance can equal or even better than it in model-based schemes such as computed torque control and feedforward nonlinear control after enough time to learn. Simulations results demonstrate the excellent performance of our scheme. The paper is a significant extension of deterministic learning theory.