基於鑑別度最佳化之遞迴式模糊類神經網路應用於分類問題

Discriminability-Optimization-Based Recurrent Fuzzy Neural Networks for Classification Problems

朱振緯(Zhen-Wei Zhu)

指導教授 ： 吳俊德

國立暨南國際大學/科技學院/電機工程學系/博士(2013年)

摘要

鑑別度在分類器的效能提升中具有很顯著的影響，為了提升鑑別度，在本篇論文提出了基於鑑別度最佳化之遞迴式模糊類神經網路 (MDRFN)，此類神經網路對於需要辨識高度混亂參數的系統具有很好的鑑別度。MDRFN將模糊系統分為兩部分，一部分為最小化分類錯誤 (MCE)，在此部分我們更新權重來使不同的分類群最大化；另一部分為最小錯誤率訓練 (MTE)，在此部分我們利用梯度下降法來更新權重，並且將cost function減少到最小以達到分類的目的。因此，MDRFN不僅增加對於時序性訊號的處理效率，並且提高了整個系統的鑑別度。此外，為了增強遞迴式類神經網路的鑑別度，我們提出了增強鑑別度之遞迴式模糊類神經網路 (EDRFNN)。相較於MDRFN的遞迴部分只考慮自我遞迴的訊息，在此網路中，利用fully connect 方法使每一個遞迴部分也具有其他遞迴的訊息。因此，此網路在分類時序性的問題中更具有效率以及鑑別度。最後， 實驗與模擬的結果顯示此論文提出的MDRFN與EDRFNN比其他的遞迴式模糊類神經網路架構如SRNFN, TRFN, SRFN 等更具有鑑別度並且在辨識的效能上更加的快速。

關鍵字

鑑別度增強 ； 遞迴式模糊類神經網路 ； 時序性參數辨識 ； 最小化分類錯誤 ； 最小錯誤訓練

並列摘要

The discriminative capability plays a significant role in determining classification performance. To increase the discriminative capability, this thesis proposes a Takagi–Sugeno(TS)-type maximizing-discriminability-based recurrent fuzzy network (MDRFN) which can classify highly confusable patterns. The proposed MDRFN considers minimum classification error (MCE) and minimum training error (MTE). In MCE, the weights are updated by maximizing the discrimination among different classes. In MTE, the parameter learning adopts the gradient–descent method to reduce the cost function. Therefore, the novelty of MDRFN is that it not only minimizes the cost function but maximizes the discriminative capability as well. Moreover, to enhance the “discriminability”, an enhanced discriminability recurrent fuzzy neural network (EDRFNN) was proposed. The feedback topology of the proposed EDRFNN is fully connected in order to handle temporal pattern behavior. It is constructed from structure and parameter learning. Simulations and comparisons with other recurrent fuzzy neural networks verify the performance of proposed recurrent fuzzy neural network under noisy conditions. In the experiments, other RFNs, including the singleton-type recurrent neural fuzzy network (SRNFN), TS-type RFN (TRFN), and simple RFN (SRFN), are compared. Analysis results indicate that the proposed MDRFN and EDRFNN exhibit excellent classification performance.

並列關鍵字

Enhanced discriminability ； Recurrent fuzzy neural network ； Temporal classification ； Minimum-classification-error ； Minimum-training-error

參考文獻

[1] J. S. Wang and C. S. G. Lee, “Self-adaptive neuro-fuzzy inference systems for classification applications,” IEEE Trans. Fuzzy Syst., vol. 10, no. 6, pp. 790–802, Dec. 2002.

[2] L. I. Kuncheva, “How good are fuzzy IF-THEN classifiers?,” IEEE Trans. Syst., Man, Cybern., B, Cybern., vol. 30, pp. 501–509, Aug. 2000.

[3] H. Ishibuchi and T. Nakashima, “Effect of rule weights in fuzzy rule-based classification systems,” IEEE Trans. Fuzzy Syst., vol. 9, pp. 506–515, Aug. 2001.

[4] C. H. Chen, C. T. Lin, and C. J. Lin, ”A functional-link-based fuzzy neural network for temperature control” in Proc. IEEE Symp. FOCI, pp. 53-58, Apr. 2007.

[5] G. Mouzouris and J. M. Mendel, “Dynamic non-singleton fuzzy logic systems for nonlinear modeling,” IEEE Trans. Fuzzy Syst., vol. 5, no. 2, pp. 199–208, May 1997.