本論文提出一種具有高計算效率之狀態空間遞迴式神經網路，我們所提出的神經網路架構為區域回授的狀態空間架構。我們所提出的神經網路主要考量在定點式數位裝置中實現。該高運算效率狀態空間結構主要基於極點靈敏度最佳化的前提下進行合成，相較於傳統的最佳化濾波器架構，每個取樣時間需要(n+1)2個乘法運算，本論文所提出的架構僅需要4n+1個乘法運算，n為狀態空間階數。在使用倒傳遞學習演算法下，我們所提出的架構在RNN架構下與傳統最佳化結構有近似的學習效果，但卻可有效大幅降低運算負擔，最後我們以數值案例驗證了本論文所提方法的有效性。
In this thesis, a high computational efficiency recurrent neural network (RNN) with state-space realizations is proposed. The proposed RNN is local feedback. We con-sider that the RNN is implemented in fixed-point digital devices. The high compu-tational efficiency state-space is synthesized based on pole sensitivity measure minimization. In contrast to the conventional optimal structures, the proposed structure requires only 4n+1 multiplications in every sample time rather than (n+1)2 multiplications in the conventional ones, where n is the order of the state-space systems. By using back propagation learning algorithm, the proposed structure is with similar performances comparing with the conventional optimal structures, but can significantly be with lower computational burden. Finally, numerical examples are performed to illustrate the effectiveness of the proposed approach.
