研究了一种基于连续型Hopfield神经网络和粒子群优化(PSO)的线性系统辨识方法.首先建立了系统的预测模型以及辨识误差函数,然后将误差函数近似为连续型Hopfield神经网络的能量函数.利用神经网络的自我演化,得到近似的辨识参数.通过引入PSO机制,缓解了Hopfield神经网络在辨识过程可能陷入局部极小值的缺陷,增强了辨识的效果.最后,仿真研究验证了所提方法的有效性.

This paper proposes a decentralized tracking control (DTC) through data-based concurrent learning for modular reconfigurable robots (MRRs) with unknown dynamics. By using the local input-output data and reference trajectories of interconnected subsystems, a neural network (NN)-based local observer is established to acquire the MRR dynamics online. Based on the adaptive dynamic programming algorithm, the local Hamilton- Jacobi-Bellman equation is solved by a local critic NN, whose weight vector is tuned by a concurrent learning-based updating law. Then, the DTC policies are obtained, and the persistence of excitation condition is removed. The tracking error of the entire closed-loop MRR system is guaranteed to be uniformly ultimately bounded by the Lyapunov's direct method. The simulation on a 2- DOF MRR system demonstrates that the proposed DTC scheme is effective.

In this paper, an event-triggered control (ETC) scheme for zero-sum game problems is proposed. To solve the Hamilton-Jacobi-Isaacs equation of the unknown nonlinear system, the adaptive dynamic programming method with critic-identifier architecture is utilized. In order to train the neural networks (NN) more efficiently and avoid manually selecting the corresponding initial weights, the particle swarm optimization is employed. In addition, the actuator is updated aperiodically under the event-triggered framework, thus, reducing the computational burden and saving communication resources to some extend. A novel triggering rule, which can guarantee the closed-loop system to be uniformly ultimately bounded, is developed through the Lyapunov method. Finally, the effectiveness of the proposed ETC scheme is demonstrated via a simulation study.