摘要
采用递归神经网络学习非线性周期运动的吸引子轨迹．网络的拓扑结构基于非线性系统的状态空间表达式,网络权值通过时序反向传播算法调整．探讨了不同样本轨迹和网络结构对递归神经网络预测性能的影响．神经网络的性能评估建立在多条测试样本轨迹的基础上,可以更为客观地评价递归神经网络预测性能．对van der Pol方程的仿真结果表明:网络的泛化能力对训练样本轨迹的依赖性较强,从不同训练轨迹上得到的递归神经网络性能差异较大;需要选择合适的递归神经网络结构参数以提高神经网络的泛化能力．
A kind of RNN（recurrent neural network） is applied to the learning of periodic attractor trajectories for nonlinear system. The network topology is based on the state-space representation, and the network parameters are optimized by the back-propagation through time algorithm. Investigations are then conducted into the model performance influenced by different training trajectories and different structure parameters. The model evaluation rule is based on multi-trajectory, which makes the investigation more objective. Simulation results from the van der Pol system show that the generalization ability is dependent on the training trajectory, different trajectories result in a significant different prediction performance; Simulation results also show that the structure parameters of the neural network should be carefully chosen so that better generalization ability can be obtained.
出处
《控制理论与应用》
EI
CAS
CSCD
北大核心
2006年第4期497-502,共6页
Control Theory & Applications
基金
国家自然科学基金资助项目(60374064).
关键词
递归神经网络
周期吸引子
泛化能力
recurrent neural network
periodic attractor
generalization ability
