library(MASS)
library(class)
library(nnet)

# number of hidden units

S=5
Decay=0.001

TrainC<-read.table("trainC.dat")
names(TrainC)<-c("x1","x2","y")

p <- as.matrix(TrainC[, -3])
tp<-TrainC$y
tpi = class.ind(tp)

xp <- seq(min(TrainC$x1), max(TrainC$x1), length = 50); np <- length(xp)
yp <- seq(min(TrainC$x2), max(TrainC$x2), length = 50)
pt <- expand.grid(x1 = xp, x2 = yp)

par(mfcol=c(2,1))

set.seed(1)
Z.nn <- nnet(p, tpi, skip=T, softmax=T, size=S, decay=Decay, maxit=1000)
Z.nn.t <- predict(Z.nn, pt ) 

zp.nn <- Z.nn.t[,1]-Z.nn.t[,2]
plot(TrainC[, 1], TrainC[, 2], xlab = "x1", ylab = "x2", col=codes(TrainC$y)+1)
title("NN TrainC")
mtext(paste("Size=",S,", Decay=",Decay))
contour(xp, yp, matrix(zp.nn, np), add = T, levels = 0, labex = 0)
   
TestC <- read.table("ExamC.dat")
names(TestC)<-c("x1","x2","y")
N <- length(TestC[,1])

xp <- seq(min(TestC$x1), max(TestC$x1), length = 50); np <- length(xp)
yp <- seq(min(TestC$x2), max(TestC$x2), length = 50)
pt <- expand.grid(x1 = xp, x2 = yp)

yhat <- predict(Z.nn, TestC[,-3] ) 

yThat <- rep('d', length(yhat[,1]))
for(i in 1:length(yhat[,1])) { if( yhat[i,1] > yhat[i,2] ) yThat[i]="a" else yThat[i]="b" }

plot(TestC[, 1], TestC[, 2], xlab = "x1", ylab = "x2", col=codes(TestC$y)+1)
title("NN ExamC")
mtext(paste("Size=",S,", Decay=",Decay))

T.nn.t <- predict(Z.nn, pt ) 
zp.nn <- T.nn.t[,1]-T.nn.t[,2]
contour(xp, yp, matrix(zp.nn, np), add = T, levels = 0, labex = 0)