Optimum Training Schemes for Recurrent Neural Network Nonlinear Filters

9616391 Olurotimi递归神经网络（RNN）是非线性动态滤波器，J. Lo已经证明它们能够收敛到信号处理的最小方差滤波器。 然而，与利用自顶向下的参数化设计来实现滤波器的传统滤波技术不同，RNN方法是一种数据驱动的综合方法。 神经网络方法由几个通用近似定理证明，这些通用近似定理确保神经网络形式在理论上足以实现这些任务。 然而，设计者和用户都熟悉的神经网络设计的一个特征是，可以构建不同的（例如权重）网络来解决相同的问题。 该项目将采用新的概念和定量结果对噪声中的RNN的行为，以解决这个问题。 PI认识到Lo的重要存在结果，并使用Olurotimi和Das的最新结果，开发了一种改良的培训措施。 由此产生的Werbos的有序导数训练方案的搜索不一定是任何最优权重集，但对于有限类的最优权重集，也增加了估计器的效率。 所提出的研究将导致的设计方案，预计将显着减少非线性RNN滤波器的设计时间。

9616391 Olurotimi Recurrent neural networks (RNN) are nonlinear dynamical filters, and it has been shown by J. Lo that they are capable of converging to the minimum variance filter for a signal process. However, unlike conventional filtering techniques that utilize top-down, parameteric design to realize the filters, the RNN approach is a data-driven synthesis approach. the neural network approach is justified by several universal approximation theorems that ensure that the neural network form is theoretically sufficient for implementing these tasks. However, one feature of neural network design familiar to designers and users alike is that may different (e.g. in weights) networks can be constructed to solve the same problem. This project will employ novel concepts and quantitative results on the behavior of RNN's in noise in order to address this problem. Recognizing the important existence results of Lo, and using recent results of Olurotimi and Das, the PI develops a modified training measure. The resulting ordered derivatives training scheme of Werbos the searches not for must any optimum weight set, but for the restricted class of optimum weight sets that also increase the estimator efficiency. The proposed research will result in s design scheme expected to significantly reduce the design time of nonlinear RNN filters.