Neural Network Dynamic Pattern Recognition Using Optimal Control

该项目的目标是：利用最优控制理论中的概念为递归神经网络设计稳健的动态模式识别算法；通过最优控制公式中揭示的可选设计选项(例如最小时间最优控制训练)来推广递归网络的导数传播算法；使用状态反馈最优控制方案来推导快速加权方案，以便在系统剧烈扰动时使用；以及对由最优控制训练方案产生的某些神经网络动态系统进行定性分析和动态行为表征。注意到最优控制理论与递归神经网络的导数传播算法之间的形式联系，本研究将进一步揭示最优控制提供的更多设计选择。因此，最优控制公式将被用来推广递归网络中的导数传播。这项研究的第一阶段将使用一种设计选项，在经过训练的动态模式识别系统中缓冲噪声的影响。第二阶段将包括最小时间最优控制，以减少递归系统的训练时间。第三阶段将提出用于缓解激进干扰的快速方案。第四阶段将对上述方案产生的一些系统进行定性分析。

The objectives of the project are: to utilize concepts from optimal control theory in designing robust dynamic pattern recognition algorithms for recurrent neural networks; to generalize derivative propagation algorithms for recurrent networks through alternative design options revealed in the optimal control formulation (One example is minimum-time optimal control training); to use state feedback optimal control schemes for deriving fast weight schemes for use during radical system disturbances; and to perform qualitative analysis and dynamic behavior characterization for certain of the neural network dynamic systems arising from the optimal control training schemes. Noting a formal relationship between optimal control theory and the derivative propagation algorithms of recurrent neural networks, this research will go further to reveal more design options offered by optimal control. Thus an optimal control formulation will be used to generalize derivative propagation in recurrent networks. The first phase of this research will use a design option that buffers against the effects of noise in trained dynamic pattern recognition systems. The second phase will incorporate minimum-time optimal control to reduce training time in recurrent systems. The third phase will propose fast schemes for use in mitigating radical disturbances. The fourth phase will perform a qualitative analysis for some of the systems resulting from the above schemes.