# Aufgabe 1

cr<-read.table("Datasets/crabs.txt",T)
library(nnet)
cr$sex<-as.factor(cr$sex)
attach(cr)
n<-nnet(sex~FL+RW+CL+CW+BD,data=cr,size=1)
table(predict(n,type="c"),sex)
summary(n)
n<-nnet(sex~FL+RW+CL+CW+BD,data=cr,skip=TRUE,size=1)
table(predict(n,type="c"),sex)
summary(n)

mkgrid <- function(x,y,points=50) {
  points<-rep(points,length.out=2)
  xv<-seq(min(x),max(x),length=points[1])
  yv<-seq(min(y),max(y),length=points[2])
  matrix(c(rep(xv,points[2]),rep(yv,each=points[1])),ncol=2)
}

g<-mkgrid(FL,RW,100)
g<-as.data.frame(g)
names(g)<-c("FL","RW")
n<-nnet(sex~FL+RW,data=cr,size=1,maxit=1000)
plot(FL,RW,col=unclass(sex))
points(g,col=predict(n,g,type="c"),pch=3)

# Aufgabe 2

# die einfachsten Loesungen fuer XOR punkte:
# - 2i-1h-1o + skip
# - 2i-2h-1o
# fuer XOR-Flaechen:
# - 2i-2h-1o + skip
# - 2i-3h-1o

xd<-data.frame(
  X=c(rnorm(100),rnorm(100)+6),
  Y=c(rnorm(50),rnorm(100)+6,rnorm(50)),
  cl=as.factor(rep(c(1,2),each=50,length.out=200)))

g<-as.data.frame(mkgrid(xd$X,xd$Y,100))
names(g)<-c("X","Y")

n<-nnet(cl~.,xd,size=1,skip=T,maxit=1000)
plot(xd$X,xd$Y,col=unclass(xd$cl))
points(g,col=predict(n,g,type="c"),pch=3)

# Xtra zum Probieren:
xd<-data.frame(
  X=c(runif(400),runif(400)+1),
  Y=c(runif(200),runif(400)+1,runif(200)),
  cl=as.factor(rep(c(1,2),each=200,length.out=800)))
g<-as.data.frame(mkgrid(xd$X,xd$Y,100))
names(g)<-c("X","Y")
n<-nnet(cl~.,xd,size=3,skip=F,maxit=1000)
plot(xd$X,xd$Y,col=unclass(xd$cl))
points(g,col=predict(n,g,type="c"),pch=3)
# und zuletzt ein ganz nettes Netz fuer die obigen Daten:
n$wts<-c(0,3,0,0,0,-3,-3,3,3,-38,-800,800,800)
# daten sind in 0..2, also 3=1.5 fuer 0..1; 800=# Faelle
plot(xd$X,xd$Y,col=unclass(xd$cl))
points(g,col=predict(n,g,type="c"),pch=3)

# Aufgabe 3

o<-read.table("Datasets/Olives/olives",T)
otr<-subset(o,o$Test=="Training")[2:10]
ott<-subset(o,o$Test=="Test")[2:10]

n<-nnet(Region~.,otr,size=3,linout=T,skip=T)
table(predict(n,type="c"),otr$Region)
table(predict(n,ott,type="c"),ott$Region)

# Aufgabe 4

cr<-read.table("Datasets/crabs.txt",T)
cr<-cbind(cr,group=factor(cr$sex+cr$sp*2-2))
i<-sample(200,50) # hier z.B. tr:test = 3:1
ic<-setdiff(1:200,i)
ctt<-cr[4:9][i,]
ctr<-cr[4:9][ic,]

n<-lapply(1:50, function(k) nnet(group~.,ctr,size=k))

etr<-unlist(lapply(n,function(x) sum(predict(x,type="c")!=ctr$group)))/150
plot(etr)
ett<-unlist(lapply(n,function(x) sum(predict(x,ctt,type="c")!=ctt$group)))/50
points(ett,col=2)
plot(ett/etr)

# Aufgabe 5

olives<-read.table("Datasets/Olives/olives",T)
olives$Area <- as.integer(olives$Area)
attach(olives)

n<-length(olives[,1])
fold <- 5
chunk <- trunc(n/fold)
nn <- chunk*fold
runs <- 20


CV <- matrix(0, runs, 1)

for(i in (1:runs)) {
   for(j in 1:fold) {
      test <- (1:chunk)*fold-fold+j
      train <- (1:nn)[-test]
      onet <- nnet(olives[3:10], Area, size=i, subset=train, 
                   decay=0.001, skip=T, linout=T, maxit=1000)
      pred <- round(predict(onet, newdata=olives[test,3:10]))
      pred[pred>9] <- 9
      pred[pred<1] <- 1
      pred<-factor(pred, levels=1:9)
      conf <- as.matrix(table(pred, Area[test]))
      CV[i] <- CV[i] + sum(conf - diag(diag(conf), 9, 9))
   }
}
plot(CV, xlab="Size", pch="o")
lines(lowess(CV ~ (1:runs)))