資源描述:
《淺議基于徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)的曲面(線)重構(gòu)仿真研究》由會(huì)員上傳分享,免費(fèi)在線閱讀,更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫(kù)。
1、廣西大學(xué)碩士學(xué)位論文基于徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)的曲面(線)重構(gòu)仿真研究姓名:楊吟冬申請(qǐng)學(xué)位級(jí)別:碩士專業(yè):計(jì)算機(jī)軟件與理論指導(dǎo)教師:周永權(quán)20070610基于徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)的曲面(線)重構(gòu)仿真研究摘要徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)以其簡(jiǎn)單的結(jié)構(gòu),優(yōu)良的全局逼近性能而引起了人們的廣泛關(guān)注。由于徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)的獨(dú)特的拓?fù)浣Y(jié)構(gòu)和訓(xùn)練方法,使得它在函數(shù)逼近和非線性系統(tǒng)預(yù)測(cè)等領(lǐng)域得到廣泛地應(yīng)用。近年來國(guó)內(nèi)外在徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)用于曲面重構(gòu)方面的研究工作比較多,普遍人們常用的是使用三個(gè)獨(dú)立的徑向基函數(shù)網(wǎng)絡(luò)分別得出曲面的三坐標(biāo)與參數(shù)之間的關(guān)系,從而間接得到曲面三坐標(biāo)之間的
2、關(guān)系,這勢(shì)必將影響到網(wǎng)絡(luò)的訓(xùn)練速度和曲面重構(gòu)的精度。本文研究如何用一個(gè)徑向基函數(shù)先直接得到擬重構(gòu)曲面的一種映射關(guān)系,這種映射關(guān)系通過網(wǎng)絡(luò)的權(quán)值和閥值來修正,其修正方法采用梯度下降算法,通過該算法對(duì)其映射關(guān)系訓(xùn)練,通過訓(xùn)練學(xué)習(xí).逐次逼近到所要擬重構(gòu)曲面.該重構(gòu)算法具有很強(qiáng)的魯棒性和較高的重構(gòu)精度.此外.本文還給出了徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)散布常數(shù)選擇的一種方法,因?yàn)樯⒉汲?shù)選得太多,易導(dǎo)致過擬合現(xiàn)象:散布常數(shù)選得太少,易導(dǎo)致曲面重構(gòu)誤差過大.最后.本文通過仿真實(shí)驗(yàn)研究散布常數(shù)對(duì)網(wǎng)絡(luò)性能的影響以及如何用徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)去用于散亂點(diǎn)的曲面重構(gòu)。關(guān)鍵詞:神經(jīng)網(wǎng)絡(luò)
3、徑向基散布常數(shù)曲面重構(gòu)RESEARCHONSI瓜FACE(LINE)RECONSTUCTBASEDONRBFNEI瓜AI,NETWORKSABSTRACTRBFNeuralNetworks’simplestructureandexcellentapproximationcapabilityarousescholars’broadattention.BecauseofRBFNeuralNetworks’specialconnectivestructureandtrainingmethod,makeituseonfunctionapproximation
4、andnon—linearforecastsystem.PresentlydomesticandoverseasresearchonhowtouseRBFNeuralNetworktoreconstructsurfaceisveryactivity.Forexample,peopleusethreeindependentlyRBFNeuralNetworkstogaintherelationshipofsurface’Sthreecoordinateandparameters,andthengaintherelationshipofthreecoor
5、dinateindirectly,butthismethodaffectnetwork’Strainingspeedprecisionofsurfacereconstruct.ThisdissertationresearchonhowtouseoneRBFNeuralNetworkgaintherelationshipofsurface'sthreecoordinatedirectly,thisrelationshipisamendedbynetworks’powervalueandvalvevalue.Themethodadoptgradientd
6、escentdarithmetic,usethisarithmetictotrainingmappingrelationship,fromtrainingandstudy,andgetreconstructivesurfacestepbystep.Thisarithmetichaveveryhighprecisionandit’Sveryrobust,Thisdissertationalsoresearchonhowtochoosecentervector,becauseifⅡchoosetoomanycentervectorswillleadove
7、l"imitate,andwillletdownit’Sgeneralizeability;Iftoofewcentervectors,thenwillnotgetenoughlearninginformationandletdownit’Sgeneralizeabilitytoo;Besidesthisdissertationviaexperimentstudyhowshapeparameteraffectnetwork’sperformanceandhowtoUSeRBFNeuralNetworktoreconstructsurfaceonuno
8、rganizeddata.KEYWORDS:NN;RBF;shapingparameter;,surface