Recherche :

Halmstad 2006
Dynamic programming and Markov chains

Jean-Philippe CHANCELIER
Version pdf de ce document
Version sans bandeaux

Table des matières

1 Finite horizon problem

Let $ (X_{n})_{n \in {\mathbb{N}}}$ be a finite Markov chain with transition matrix $ M^{(n)}$. We want to recursively compute :

$\displaystyle v_n(x) = {\mathbb{E}} \left[ \sum_{k=n}^{T-1}\frac{1}{(1+r)^{k-n}} c_k(X_{k}) +\frac{1}{(1+r)^{T-n}}f(T,X_{T}) \vert X_{n} = x \right], $

    

Question 1   Write a first routine which returns a random stochastic matrix $ M$ of size NxN. In scilab there is a function called genmarkov that you can also check if you want. 



  
function M=my_gen_markov(n); 
  P=rand(n,n,'u');
  y=sum(P,'c'); y = ones(y)./y ; y=y*ones(1,N);
  M=y.* P;
endfuncion

    

Question 2   Choose a state size, generate a stochastic matrix $ M$, then generate and plot some samples of the homogeneous Markov chain with states in $ [1,N]$ described by $ M$ (using grand in Scilab).



  
n=10;
M=my_gen_markov(n)
// generate and draw some typical samples 
T=4; // horizon time 
m=20; // number of samples 
// generate the m-samples 
// each sample is of length T
Y=grand(T,'markov',M,ones(m,1));
// just add the initial state to each sample
Y=[ones(m,1),Y]; 
// you can check the size of Y with size 
size(Y)
// draw the first sample 
plot(Y(1,:))
// draw the second sample 
plot(Y(2,:))

    

Question 3   Choosing an instantaneous cost $ c$ and a final cost $ f$ recusively compute the value function $ v^n(x)$ and draw the result as a surface (i.e $ v(t,x)$). The computation can be stored in a matrix. 



  
function y=c(x); y=x; endfunction 
function y=f(x);K=2; y=max(x-K,0); endfunction 

states=(1:n)' // possible states 
Cv=c(states) // vector of c possible values 
fv=f(states) // vector of f possible values 
// here you should check that Cv and fv are column vectors

r=0.05; // a discount factor 
// I compute the value function and store the 
// results in a matrix V 
V=zeros(n,T);
V(:,T) = fv;
for i=T-1:-1:1 
  V(:,i) = (Cv + M*V(:,i+1))/(1+r);
end
// plot the surface 

plot3d(1:n,1:T,V)

    

Question 4   Using samples of the markov chain for a given starting state evaluate by Monte Carlo the cost function $ V(1,1)$ and compare the results with previous question. 



  
// test by monte carlo 
// Compute V(1,1) by monte carlo 
m=10000;
X0=ones(m,1);
X=grand(T-1,'markov',M,X0);
X=[X0,X];

// fix n
// Approximate the cost by Monte Carlo 

n=1;
Cm= mean((1/(1+r).^(T-n))*f(X(:,T)));
for i=(T-1):-1:n;
  Cm = mean((1/(1+r).^(i-n))*c(X(:,i)))+ Cm;
end

// test 

Cm -V(1,1)