Recurrent Neural Network With Scheduled Varying Gain for Solving Time-Varying QP

Recurrent neural networks, especially zeroing neural networks (ZNNs) are massively deployed for solving the time-varying quadratic programming problem. Varying-gain ZNNs (VGZNNs) become attractive in recent years owing to the superior convergence performance in comparison with fixed-gain ZNN models. However, the variant gains of most existing VGZNN models tend to infinity as the time variable tends to infinity, causing the VGZNN models improper for long-period deployment. To tackle this problem, we propose a novel scheduled VGZNN (SVGZNN) model with a gain schedule mechanism. With this mechanism, the varying gain is always bounded and we can readily adjust the gain according to the gain constraint and the desirable gain for long-period deployment. Theoretical analysis on convergence properties of the proposed SVGZNN is conducted. The SVGZNN model can globally converge in finite time. Simulation results are also discussed.