Title: 使用健全學習法則的多項式類神經網路等化器
Polynomial-Perceptron Based Adaptive Equalizer with a Robust Liarning Algorithm
Authors: 張清濠
Chang,Ching-Haur
魏哲和, 蕭師基
Che-Ho Wei, Sammy Siu
電子研究所
Keywords: 最佳判別界線;多項式類神經網路判別式迴授等化器;非線性判別界線.;Optimal decision bounddary;feedback equalizer;robust learning.
Issue Date: 1995
Abstract: 本文研究非線性等化器之必要性及設計方法並提出一種新型非線性等化器
研究範圍包括下列兩個方面:第一,提出一個使用單層多項式類神經網路
的判別式迴授等化器,第二,提出一個使用非平方差誤差準則並適用於上
述等化器的學習法則。這種等化器結構能夠發揮判別式迴授等化器的優點
並利用單層類神網路結構達到所需的非線性,因此在引用隨機梯度濱算法
則時較之多層結構容易的多,經由詳細的數學分析,我們證實使用次冪小
於二的非平方差誤差準則對此新型等化器的誤差大小確實有健全的處理方
式,正因如此,這種演算法可以達到很好的收斂性,其中關於次冪小於一
所可能產生的數值問題,我們提出兩種解決方法並經由理論驗證其有效性
,計算機模擬結果顯示這種新型等化器在收斂速率及誤碼率上均有很好的
性能,而且也驗證了使用次冪小於二的學習法則對這種新型等化器的優異
性,尤其值得一提的是,這種新型等化器在解決嚴重非線性失真問題時能
夠達到幾乎與貝氏準則相同的性能,而且在某些情況下,它的性能甚至於
可以接近最大相似性序列估計準則所達到的性能,我們在模擬貝氏準則與
最大相似性序列估計準則時,使用已知的信道響應參數,因此所得數據足
可做為性能指標。 In this dissertation, a new nonlinear
equalizer is addressed
in the following two respects. First, a single-layer non-
linear decision feedback equalizer (DFE) equipped with
polynomial-perceptron model of nonlinearities is developed.
Second, an lp-norm based learning algorithm suitable for the
addresseed structure is investigated. The structure exerts the
benefit of using a DFE and achieves the required non-
linearities in a single-layer net. This is advantageous since
it is much easier to train by a stochastic gradient algorithm .
It is shown that the algorithm using lp-norm error cir- terion
with p<2 is robust to deal with the error for the new equalizer
and hence better convergence property is obtained. Detailed
performance analysis with a consideration on the possible
numerical problem for p<1 is performed. Computer simulations
show that the new equalizer is satisfactory in both convergence
rate and bit error rate(BER) performance. Also, our scheme is
shown capable of achieving the performanc offered by a minimum
BER equalizer. In some case, a per- formance quite close to the
maximum-likelihood sequence estimation(MLSE)criterion can also
be attained.
URI: http://140.113.39.130/cdrfb3/record/nctu/#NT840430130
http://hdl.handle.net/11536/60740
Appears in Collections:Thesis