針對EKF在非線性系統中應用的不足，提出了一個基於取樣策略的改良KF，即無跡KF（UKF）。該策略看起來像是一種機器學習，因此我們採用了一種基於神經網路（NN）以取代 EKF 的演算法。 EKF 和 NN 的數值實作皆在馬達和簡單線性電路的範例中完成。根據計算實驗的結果，神經網路在非線性系統的狀態估計中似乎比 EKF 更可靠。

According to shortcomings from application of EKF for a nonlinear system, a modified KF based on sampling strategy, named unscented KF (UKF), is proposed. The strategy seems like a kind of machine learning, we hence adopt an algorithm based on neural network (NN) instead of EKF. Numerical Implementation for both EKF and NN is accomplished on examples of motors and simple linear circuits. Based on computation experiments, better than EKF, NN seems more reliable in state estimation for a nonlinear system.

[1]Eric A. Wan and Rudolph van der Merwe, The Unscented Kalman Filter for Nonlinear Estimation, IEEE, 06 August 2002.
[2] Simon J. Julier, Member, IEEE, and Jeffrey K. Uhlmann, Member, Unscented Filtering and Nonlinear Estimation, IEEE, vol. 92, NO. 3, March 2004.
[3] Oppenheim, Alan V./ Willsky, Alan S./ Nawab, Syed Hamid, Signals & Systems, Prentice Hall, 2008-03-31.
[4] Hasan A. Bjaili1 · Muhammad Moinuddin2 · Ali M. Rushdi, A State-Space Backpropagation Algorithm for Nonlinear Estimation, Circuits, Systems, and Signal Processing, 11 January 2019.
[5] Charles Alexander, Matthew Sadiku, Fundamentals of Electric Circuits, Tata McGrawhill India, 2013
[6]程登湖, 機器人學：多軸旋翼機 Lec11, NYCU OCW 機械工程學系, 2022/5/19.
[7] Simon Haykin, Neural Networks and Learning Machines, 2008/11/18.
[8] 林長鋆,【深度學習實作】課程 1.2 模型訓練與觸發函數, 2022/8/1.
[9] Marius Hobbhahn, Jaime Sevilla, What’s the Backward-Forward FLOP Ratio for Neural Networks, Epoch AI, 2021/12/13
[10] 吳尚鴻, 深度學習課程 ML-10:Neural Networks NN-Basics, 2017/8/30.
[11] 伊藤 真, Ｐｙｔｈｏｎで動かして学ぶ！あたらしい機械学習の教科書, 日本出版株式会社, 2022/07.
[12] 洪錦魁, 演算法：最強彩色圖鑑 + Python程式實作 王者歸來, 深智數位, 2020/04/02.