一種神經(jīng)網(wǎng)絡(luò)穩(wěn)健估計(jì)方法的推廣性研究
GENERALIZATION STUDY FOR A ROBUST ESTIMATION METHOD OF NEURAL NETS
-
摘要: 本文根據(jù)統(tǒng)計(jì)學(xué)的穩(wěn)健性原理,將柯西(Cauchy)函數(shù)作為新的神經(jīng)網(wǎng)絡(luò)目標(biāo)函數(shù)。在網(wǎng)絡(luò)參數(shù)相同的前提下,利用傳統(tǒng)的均方目標(biāo)函數(shù)和新的柯西目標(biāo)函數(shù)對(duì)BP網(wǎng)絡(luò)分別進(jìn)行訓(xùn)練后,加入小噪聲及異常值(Outliers)干擾對(duì)該網(wǎng)絡(luò)進(jìn)行測(cè)試。結(jié)果表明,具有穩(wěn)健性目標(biāo)函數(shù)的網(wǎng)絡(luò)不但有更快的收斂速度,而且對(duì)異常值有更好的抵抗能力。
-
關(guān)鍵詞:
- 穩(wěn)健誤差估計(jì)器; BP算法; 目標(biāo)函數(shù); 異常值
Abstract: In this paper, the Cauchy function is taken as a new target function of neural network accordings to the robustness theorem of statistics. Under the same network parameter conditions the BP net is trained using both mean squresand Cauchy target function firstly, then the net is tested by data sets including small Gaussian noises and outliers separately. Simulation results indicate that the network has both faster convergence speed and better performance against outliers after learning with robust target function. -
Huber P J. Robust estimation of a location parameter[J].The Annals of Mathematical Statistics.1964, 35:73-101[2]Andrews D A. Robust method for mutiple linear regression[J].Technometrics.1974, 16:523-531[3]Liano K. Robust error measure for supervised neural network learning with outliers. IEEE Trans. on Neural Networks, 1996, NN-7(1): 244-250.[4]Chen D S, Jain R C. A robust back propagation learning algorithm for function appoximation. IEEE Trans. on Neural Networks, 1994, NN-5(3): 467-469.[5]Oja E, Wang L. Robust fitting by nonlinear neural units[J].Neural Networks.1996, 9(3):435-444[6]Humpert B K. Improving back propagation with a new error function. Neural Nerworks, 1994, 7(R)- 1101-1149.[7]陳希孺, 王松桂. 近代實(shí)用回歸分析. 南寧:廣西人民出版社, 1984: 301-321.[8]廖曉峰, 劉光遠(yuǎn), 虞厥邦.幾種誤差估計(jì)器的穩(wěn)健BP:理論與算法.信號(hào)處理,1997, 13(3): 235-240. -
計(jì)量
- 文章訪問數(shù): 2200
- HTML全文瀏覽量: 108
- PDF下載量: 408
- 被引次數(shù): 0
下載:
