1 簡(jiǎn)介

人工神經(jīng)網(wǎng)絡(luò)的最大缺點(diǎn)是訓(xùn)練時(shí)間太長(zhǎng)從而限制其實(shí)時(shí)應(yīng)用范圍,近年來(lái),極限學(xué)習(xí)機(jī)(Extreme Learning Machine, ELM)的提出使得前饋神經(jīng)網(wǎng)絡(luò)的訓(xùn)練時(shí)間大大縮短,然而當(dāng)原始數(shù)據(jù)混雜入大量噪聲變量時(shí),或者當(dāng)輸入數(shù)據(jù)維度非常高時(shí),極限學(xué)習(xí)機(jī)算法的綜合性能會(huì)受到很大的影響.深度學(xué)習(xí)算法的核心是特征映射,它能夠摒除原始數(shù)據(jù)中的噪聲,并且當(dāng)向低維度空間進(jìn)行映射時(shí),能夠很好的起到對(duì)數(shù)據(jù)降維的作用,因此我們思考利用深度學(xué)習(xí)的優(yōu)勢(shì)特性來(lái)彌補(bǔ)極限學(xué)習(xí)機(jī)的弱勢(shì)特性從而改善極限學(xué)習(xí)機(jī)的性能.為了進(jìn)一步提升DELM預(yù)測(cè)精度，本文采用麻雀搜索算法進(jìn)一步優(yōu)化DELM超參數(shù)，仿真結(jié)果表明，改進(jìn)算法的預(yù)測(cè)精度更高。

2 部分代碼

clc;clear all;close all;[f p]=uigetfile('*');X=importdata([p f]);data=X.data;data2=data(:,1:end-1);class=data(:,end);data1=knnimpute(data2);%%%%%%%%Feature selection FSL=0;FSU=1;D=size(data1,2);for i=1:10FS(i,:)=FSL+randi([0 1],[1 D])*(FSU-FSL);tryfit(i)=fitness(data1,class,FS(i,:));catch ? ?fit(i)=1; ? ?continue;endendind=find(fit==min(fit));FSnew=FS(ind,:);pdp=0.1;row=1.204;V=5.25;S=0.0154;cd=0.6;CL=0.7;hg=1;sf=18;Gc=1.9;D1=1/(2*row*V.^2*S*cd);L=1/(2*row*V.^2*S*CL);tanpi=D1/L;dg=hg/(tanpi*sf);aa=randi([1 length(ind)]);iter=1;maxiter=2;while(iter<maxiter)for i=1:10if(rand>=pdp) ? ?FS(i,:)=round(FS(i,:)+(dg*Gc*abs(FSnew(1,:)-FS(i,:))));else ? FS(i,:)=FSL+randi([0 1],[1 D])*(FSU-FSL); endFh=FS;fit1(i)=fitness(data1,class,FS(i,:));ind1=find(fit1==min(fit1));FSnew1=FS(ind1,:);if(rand>pdp) ? ?FS(i,:)=round(FS(i,:)+(dg*Gc*abs(FSnew(aa,:)-FS(i,:))));else ? FS(i,:)=FSL+randi([0 1],[1 D])*(FSU-FSL); endFa=FS;fit2(i)=fitness(data1,class,FS(i,:));ind2=find(fit2==min(fit2));FSnew2=FS(ind2,:);endSc=sqrt(sum(abs(Fh-Fa)).^2);Smin=(10*exp(-6))/(365).^(iter/(maxiter/2.5));if(Sc<Smin) ? ?season=summer; ? ?for i=1:10 ? ? ? ?FS(i,:)=FSL+levy(1,D,1.5)*(FSU-FSL); ? ?end else ? ?season=winter; ? ?break;end%%%Searching methodfit3(i)=fitness(data1,class,FS(i,:));ind3=find(fit3==min(fit3));final=abs(round([Fh(ind1,:);Fa(ind2,:);FS(ind3,:)]));for i=1:size(final,1) ? ?fitt(i)=fitness(data1,class,final(i,:));endbest(iter)=min(fitt);[ff inn]=min(fitt);bestfeat(iter,:)=final(inn,:);pdp=best(iter);iter=iter+1;endsel=find(bestfeat(end,:));disp('Selected Features');disp(sel)dataA =data2(:,sel); ?% some test datap = .7 ; ? ? % proportion of rows to select for trainingN = size(dataA,1); ?% total number of rows tf = false(N,1); ? % create logical index vectortf(1:round(p*N)) = true; ? tf = tf(randperm(N)); ? % randomise orderdataTraining = dataA(tf,:);labeltraining=class(tf);dataTesting = dataA(~tf,:);labeltesting=class(~tf);disp('Training feature size');disp(length(dataTraining))disp('Testing feature size');disp(length(dataTesting))svt=svmtrain(dataTraining,labeltraining);out1=svmclassify(svt,dataTesting);mdl=fitcknn(dataTraining,labeltraining);out2=predict(mdl,dataTesting);%%%%%%% ?NB %%%%%%%%mdl=fitcensemble(dataTraining,labeltraining);out3=predict(mdl,dataTesting);tp=length(find(out3==labeltesting));msgbox([{['Out of ',num2str(length(out3))]},{[num2str(tp),'are correctly classified']}])delete(gcp('nocreate'))disp('%%%%%%%% ?KNN %%%%%%%%%%%%%%')[EVAL CF] = Evaluate(out2,labeltesting);disp('Accuracy (%)');disp(EVAL(1)*100);disp('Precision (%)');disp(EVAL(4)*100);disp('Recall (%)');disp(EVAL(5)*100);disp('Fmeasure (%)');disp(EVAL(6)*100);disp('True Positive');disp(CF(1))disp('True Negative');disp(CF(2))disp('False Positive');disp(CF(3))disp('False Negative');disp(CF(4))disp('%%%%%%%% ?SVM %%%%%%%%%%%%%%')[EVAL3 CF] = Evaluate(out1,labeltesting);disp('Accuracy (%)');disp(EVAL3(1)*100);disp('Precision (%)');disp(EVAL3(4)*100);disp('Recall (%)');disp(EVAL3(5)*100);disp('Fmeasure (%)');disp(EVAL3(6)*100); disp('True Positive');disp(CF(1))disp('True Negative');disp(CF(2))disp('False Positive');disp(CF(3))disp('False Negative');disp(CF(4))disp('%%%%%% ?NB %%%%%%%%%%%%%%')[EVAL2 CF] = Evaluate(out3,labeltesting);disp('Accuracy (%)');disp(EVAL2(1)*100);disp('Precision (%)');disp(EVAL2(4)*100);disp('Recall (%)');disp(EVAL2(5)*100);disp('Fmeasure (%)');disp(EVAL2(6)*100);disp('True Positive');disp(CF(1))disp('True Negative');disp(CF(2))disp('False Positive');disp(CF(3))disp('False Negative');disp(CF(4))

3 仿真結(jié)果

4 參考文獻(xiàn)

[1]馬萌萌. 基于深度學(xué)習(xí)的極限學(xué)習(xí)機(jī)算法研究[D]. 中國(guó)海洋大學(xué), 2015.

博主簡(jiǎn)介：擅長(zhǎng)智能優(yōu)化算法、神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)、信號(hào)處理、元胞自動(dòng)機(jī)、圖像處理、路徑規(guī)劃、無(wú)人機(jī)等多種領(lǐng)域的Matlab仿真，相關(guān)matlab代碼問(wèn)題可私信交流。

部分理論引用網(wǎng)絡(luò)文獻(xiàn)，若有侵權(quán)聯(lián)系博主刪除。