資源描述:
《近鄰傳播聚類算法的rbf隱含層節(jié)點優(yōu)化》由會員上傳分享,免費在線閱讀,更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、近鄰傳播聚類算法的RBF隱含層節(jié)點優(yōu)化 摘要:傳統(tǒng)的RBF神經(jīng)網(wǎng)絡(luò)預(yù)測精度會由于隨機(jī)選取隱含層中心節(jié)點不合適而導(dǎo)致算法效率低下和數(shù)值病態(tài),為了提高RBF神經(jīng)網(wǎng)絡(luò)的效率,提出了一種用近鄰傳播AP聚類算法改進(jìn)RBF神經(jīng)網(wǎng)絡(luò)的方法,并介紹了該方法的原理及建模步驟。由于采用的AP聚類算法屬于自適應(yīng)聚類學(xué)習(xí)算法,無需事先給定隱含層中心節(jié)點的個數(shù),能夠適用于不具有先驗信息的預(yù)測。首先,利用AP算法根據(jù)訓(xùn)練樣本的信息進(jìn)行聚類迭代,從而確定RBF神經(jīng)網(wǎng)絡(luò)中隱含層的中心節(jié)點和節(jié)點數(shù)值,解決了RBF網(wǎng)絡(luò)的中心取值問題。然后,把所有輸入數(shù)據(jù)代入基于AP聚類算法優(yōu)化的RBF神經(jīng)網(wǎng)絡(luò)中進(jìn)行預(yù)測。由于
2、AP算法無需預(yù)先指定聚類數(shù)目,所提方案能提高網(wǎng)絡(luò)的學(xué)習(xí)精度和訓(xùn)練速度,利用所提優(yōu)化方案對正弦函數(shù)進(jìn)行逼近的仿真實驗,結(jié)果表明該方案的逼近誤差僅為0.0055,在0.3噪聲下能保持較好的預(yù)測精度?! £P(guān)鍵詞:徑向基函數(shù)神經(jīng)網(wǎng)絡(luò);近鄰傳播聚類算法;隱含層;逼近誤差 中圖分類號:TN711?34;TP398.1文獻(xiàn)標(biāo)識碼:A文章編號:1004?373X(2016)19?0016?04 Abstract:Thepredictionaccuracyofthetraditionalradialbasisfunction(RBF)neuralnetworkmayresultinlowe
3、ralgorithmefficiencyandpathologicalnumericalvalueduetothe7inappropriaterandomselectionofthehiddenlayercenternode,toimprovetheefficiencyofRBFneuralnetwork,amethodofusingaffinitypropagation(AP)clusteringalgorithmtoimproveRBFneuralnetworkisproposed.Theprincipleandmodelingstepsofthemethodareint
4、roduced.SincetheadoptedAPclusteringalgorithmbelongstotheself?adaptingclusteringlearningalgorithm,itneedn′tpredefinethenumbersofthehiddenlayercenternodes,andisappliedtopredictionwithouttranscendentalinformation.TheAPalgorithmisusedforclusteringiterationaccordingtheinformationoftrainingsample
5、,soastodeterminethecenternodeandnodenumericalvalueofhiddenlayerinRBFneuralnetwork,andsolvethecenterdereferencingproblemofRBFnetwork.Afterthat,allinputdataistakeninRBFneuralnetworkbasedonAPclusteringalgorithmforprediction.SincetheuseofAPalgorithmneedn′tpredefinetheclusteringnumbers,thepropos
6、edschemecanimprovethelearningaccuracyandtrainingspeedoftheRBFneuralnetwork.Theapproximatesimulationexperimentwasperformedforsinefunctionwiththeproposedoptimizationscheme.Theresultsshowthattheapproximateerroroftheproposedschemeisonly0.0055,andcankeepgoodpredictionaccuracyunderthenoiseof0.3.
7、 Keywords:radialbasisfunctionneuralnetwork;affinity7propagationclusteringalgorithm;hiddenlayer;approximateerror 0引言 RBF(RadialBasisFunction)網(wǎng)絡(luò)是一種單隱含層前饋神經(jīng)網(wǎng)絡(luò),其基本思想是在隱含層內(nèi)基函數(shù)的作用下,將輸入信息的不可分矢量變換到高維可分空間[1?3]。RBF網(wǎng)絡(luò)結(jié)構(gòu)簡單而且具備非線性逼近能力,收斂速度快。RBF網(wǎng)絡(luò)已經(jīng)廣泛應(yīng)用于函數(shù)逼近、模式識別、信號