R語言實(shí)現(xiàn)人工神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)

R語言中很多包（package）關(guān)于神經(jīng)網(wǎng)絡(luò)，例如nnet、AMORE、neuralnet以及RSNNS。nnet提供了最常見的前饋反向傳播神經(jīng)網(wǎng)絡(luò)算法。AMORE包則更進(jìn)一步提供了更為豐富的控制參數(shù)，并可以增加多個(gè)隱藏層。neuralnet包的改進(jìn)在于提供了彈性反向傳播算法和更多的激活函數(shù)形式。RSNNS則是連接R和SNNS的工具，在R中即可直接調(diào)用SNNS的函數(shù)命令，在這方面有了極大的擴(kuò)充。本文使用AMORE包來實(shí)現(xiàn)神經(jīng)網(wǎng)絡(luò)模型?？墒褂妹頸nstall.packages("AMORE")進(jìn)行包的安裝。

R語言神經(jīng)網(wǎng)絡(luò)代碼示例如下：

library(AMORE)  
      
    #輸入一個(gè)11*10的矩陣，前8行用來訓(xùn)練，后3行用來預(yù)測(cè)  
    p<-matrix(c(6977.93,24647,11356.6,9772.5,1496.92,4279.65,89.84,95.97,9194,0.6068,  
                7973.37,28534,13469.77,11585.82,1618.27,5271.991,100.28,111.16,9442,0.63,  
                9294.26,33272,16004.61,14076.83,1707.98,6341.86,117.78,130.22,9660,0.6314,  
                10868.67,37638,18502.2,16321.46,1790.97,6849.688,134.77,125.56,9893,0.6337,  
                12933.12,39436,19419.7,18052.59,1855.73,6110.941,86.04,119.81,10130,0.634,  
                15623.7,44736,23014.53,20711.55,1948.06,7848.961,151.59,187.08,10441,0.6618,  
                17069.2,50807,26447.38,24097.7,2006.92,9134.673,177.79,202.12,10505,0.665,  
                18751.47,54095,27700.97,26519.69,2037.88,9840.205,195.18,282.05,10594,0.674,  
                21169.7,60633.82,31941.45,29569.92,2211.6665,11221.01,205.5601,329.4234,10986.79,0.684065,  
                23716.17,66750.29,35562.93,32993.75,2317.9223,12486.77,220.3005,398.7751,11245.69,0.694706,  
                26469.74,73292.95,39458.17,36680.63,2428.5869,13849.68,235.0408,477.4204,11515.33,0.706087),11,10,byrow=T)  
      
    #對(duì)輸入矩陣進(jìn)行歸一化處理（0到1）  
    b1=(p[,1]-min(p[,1]))/(max(p[,1])-min(p[,1]))    
    b2=(p[,2]-min(p[,2]))/(max(p[,2])-min(p[,2]))    
    b3=(p[,3]-min(p[,3]))/(max(p[,3])-min(p[,3]))    
    b4=(p[,4]-min(p[,4]))/(max(p[,4])-min(p[,4]))   
    b5=(p[,5]-min(p[,5]))/(max(p[,5])-min(p[,5]))    
    b6=(p[,6]-min(p[,6]))/(max(p[,6])-min(p[,6]))    
    b7=(p[,7]-min(p[,7]))/(max(p[,7])-min(p[,7]))    
    b8=(p[,8]-min(p[,8]))/(max(p[,8])-min(p[,8]))   
    b9=(p[,9]-min(p[,9]))/(max(p[,9])-min(p[,9]))    
    b10=(p[,10]-min(p[,10]))/(max(p[,10])-min(p[,10]))    
    p0=cbind(b1,b2,b3,b4,b5,b6,b7,b8,b9,b10)#歸一化后的數(shù)據(jù)放入矩陣中   
      
    #對(duì)應(yīng)矩陣前8行的測(cè)試結(jié)果集  
    t<-c(2673.5356,2991.0529,3393.0057,3504.8229,3609.4029,4060.1257,4399.0168,4619.4102)  
      
    #第9行的實(shí)際結(jié)果  
    t9=4830.1315  
      
    #測(cè)試結(jié)果歸一化  
    t0=(t-min(t))/(max(t)-min(t))  
      
    alter=1  
    count=0  
      
    #訓(xùn)練的結(jié)果測(cè)試第9行若誤差在3%之內(nèi)或者循環(huán)20次結(jié)束  
    while(abs(alter)>0.03 && count<20){  
        
        #訓(xùn)練網(wǎng)絡(luò)，n.neurons表示輸入的參數(shù)，以及隱藏層個(gè)數(shù)，及輸出結(jié)果  
      net<-newff(n.neurons = c(10,10,2,1),learning.rate.global=1e-4, momentum.global=0.05,error.criterium="LMS", Stao=NA, hidden.layer="tansig", output.layer="purelin", method="ADAPTgdwm")  
      #<span style="line-height: 27.2px; font-family: 'Helvetica Neue', Helvetica, Tahoma, Arial, STXihei, 'Microsoft YaHei', 微軟雅黑, sans-serif;">p0[1:8,]表示輸入，t0[1:8]表示輸出，show.step表示循環(huán)次數(shù)，n.shows表示滿足結(jié)果的報(bào)告次數(shù)</span>  
      result<-train(net,p0[1:8,],t0[1:8],error.criterium="LMS", report=TRUE, show.step=10000, n.shows=5)  
      #測(cè)試第9行到11行  
      y<-sim(result$net,p0[9:11,])  
      #反歸一化  
      y<-y*t[8]  
      #用第9行來測(cè)試訓(xùn)練誤差，滿足訓(xùn)練誤差結(jié)束  
      alter=(y[1]-t9)/t9  
        
      count=count+1;  
        
    }  
      
    count  
    #輸出第9行到11行預(yù)測(cè)的值  
    y  
    #作圖  
    x0<-c(2013,2014,2015)  
    plot(x0,y,col = "blue",pch = "+") 

注：每一行都具有實(shí)際含義，代表每一年的參數(shù)指標(biāo)，預(yù)測(cè)一個(gè)值