# przykład 6.7

options(OutDec=",")
library("tree")
library("ElemStatLearn")
library("cluster")
library("mlbench")
library("rpart")

mvote0<-function(x)
{
t<-table(x)
ll<-length(t)
tm<-as.matrix(t)
m<-rownames(tm)[which.max(t)]
if (sum(t==max(t))>1)
{
c<-which(t==max(t))
m<-rownames(tm)[sample(c,1)]
}
return(m)
}


data(Satellite)
mo<-nrow(Satellite)

ks<-seq(5,100,by=5)
blad.sat.c<-matrix(nrow=length(ks),ncol=100)

for (j in 1: 100)
{
print(j)
ucz <- sample(1:mo, size = 4290, replace = FALSE) 
sat.ucz<-Satellite[ucz,]
sat.test<-Satellite[-ucz,]

y.test<-sat.test$classes

y<-sat.ucz$classes
y<-as.data.frame(y)
sat.ucz<-sat.ucz[,-ncol(sat.ucz)]

m<-ncol(sat.ucz)

for (k in seq(along=ks))
{
lmod<-ks[k] # liczba modeli pojedynczych w ensenble

yklas<-NULL

for (i in 1:lmod)
{ # drzewa
zmienne <- sample(1:m, size = round(m/2), replace = FALSE) 
zb.ucz<-sat.ucz[,zmienne]
zb.ucz<-cbind(y,zb.ucz)
sat.tr<-rpart(y~.,zb.ucz)

Z.test<-predict(sat.tr,sat.test,type ="class")
yklas<-cbind(yklas,as.matrix(Z.test))
}

mm<-nrow(yklas)
ypred<-numeric(length=mm)
for (i in 1:mm) ypred[i]<-mvote0(yklas[i,])

blad.sat.c[k,j]<-1-sum(y.test==ypred)/length(ypred)
}
}

# Rysunek 6.6

sat.rsm<-numeric(length=20)
for (k in 1:20) sat.rsm[k]<-mean(blad.sat.c[k,])
print(sat.rsm)

plot(ks,sat.rsm,type="n",xlab="Liczba modeli bazowych",ylab="Błąd predykcji",main="Zbiór Satellite")
lines(ks,sat.rsm,type="l",lty=1)
points(ks,sat.rsm,pch=19)