Compound gradient vector based neural networks for real-time control

An improved compound gradient vector based a NN online training weight update scheme is proposed in this paper. The convergent analysis indicates that because the compound gradient vector is employed during the weight update, the convergent speed of the presented algorithm is faster than the BP algorithm. In this scheme an adaptive learning factor is introduced, in which the global convergence is obtained, and the convergent procedure on plateau and flat bottom area can speed up. Simulations have been conducted and the results demonstrate that the satisfactory convergent performance and strong robustness are obtained using the improved compound gradient vector NN online learning scheme for real time control.