north=c(1:3,15,1:10)
east=2:15;east[ c(3,7,11)]=c(3,7,11)
south=c(5:15,12:14)
west=c(15,1:13);west[ c(4,8,12)]=c(4,8,12)
trM=cbind(north,east,south,west)
###

nRep=1000; gamma=1;
P = 0.25
R = -1
V = rep(0,15)
for (i_rep in 1:nRep) {
  V.old=V
  for (i_state in 1:14) {
    V[i_state]=sum(P*(R+gamma*V.old[trM[i_state,]]))
  }
}
matrix(c(0,V),4,4)

#   ch4.1 policy evaluation 
########################################
# example 4.1 - bug fixed on 2017.03.21
# defining deterministic trans. matrix
north=c(1:3,15,1:10)
east=2:15;east[ c(3,7,11)]=c(3,7,11)
south=c(5:15,12:14)
west=c(15,1:13);west[ c(4,8,12)]=c(4,8,12)
trM=cbind(north,east,south,west)
 
# defining Reward & trans. prob.
R=-1;P=0.25;
 
# iterative policy evaluation
delta=1; gamma=1; tol=1e-8; V=rep(0,15);
while (delta>tol) {
 delta=0;
  for (i_state in 1:14) {
    v=V[i_state]
    V[i_state]=sum(P*(R+gamma*V[trM[i_state,]]))
    delta=max(delta,abs(v-V[i_state]));
  }
}
 
# result
> matrix(c(0,V),nrow=4)
     [,1] [,2] [,3] [,4]
[1,]    0  -14  -20  -22
[2,]  -14  -18  -20  -20
[3,]  -20  -20  -18  -14
[4,]  -22  -20  -14    0
 
#####################################
#   ch4.3 Policy Improvement
#   easier one 
#####################################
north=c(1:3,15,1:10)
east=2:15;east[ c(3,7,11)]=c(3,7,11)
south=c(5:15,12:14)
west=c(15,1:13);west[ c(4,8,12)]=c(4,8,12)
trM=cbind(north,east,south,west)
R=-1;V=rep(0,15);
delta=1; gamma=1; tol=1e-5; 
bestP=sample(1:4,14,replace=T)
stable=F;counter=0
while (stable==F){
  counter=counter+1
  Vmat=matrix(V[trM],ncol=4)
  Ppolicy=t(apply(Vmat,1,function(x) exp(10*x)/sum(exp(10*x))))
# iterative policy evaluation
  while (delta>tol) {
    delta=0;
    for (i_state in 1:14) {
      v=V[i_state]
      V[i_state]=sum(Ppolicy[i_state]*(R+gamma*V[trM[i_state,]]))
      delta=max(delta,abs(v-V[i_state]))
    }
  }
# policy improvement
  stable=F
  for (i_state in 1:14) {
    b=bestP[i_state]
    bestP[i_state]=which.max(V[trM[i_state,]])
    ifelse((bestP[i_state]==b),stable<-T,stable<-F)
  }
}
 
 
#####################################
#   ch4.4 Value Iteration 
#####################################
# example 4.3
V=c(rep(0,100),1);V.hist=c()
p=c(0.4,0.6);
gamma=1;delta=1; tol=1e-20
max.a=rep(0,101)
while (delta>tol) {
  delta=0;
  for (i_state in 1:99) {
    v=V[i_state+1]
    temp=matrix(0,nrow=1,ncol=i_state)
    for (i_action in 1:i_state) {
       temp[i_action]=sum(p*(gamma*c(V[(min(i_state+i_action,100)+1)],
         V[(max(i_state-i_action,0)+1)])))
     }
    V[i_state+1]=max(temp)
    max.a[i_state+1]=which.max(round(temp,8))
    delta=max(delta,abs(v-V[i_state+1]))
  }
  V.hist=rbind(V.hist,V)
}
# plotting results
par(mfrow=c(1,2))
plot(V.hist[1,],type='l',lwd=2,xlab="Capital",ylab="Value Estimates",col='red')
lines(V.hist[2,],lwd=2,col='blue')
lines(V.hist[3,],lwd=2,col='green')
lines(V.hist[32,],lwd=2,col='black')
legend("topleft",c("sweep 1","sweep 2","sweep 3", "sweep 32"),
 col=c("red","blue","green","black"),lwd=2)
barplot(max.a,xlab="Capital",ylab="Final Policy",col="white")