1
$\begingroup$

What is the optimal policy for this MDP?

enter image description here

$\endgroup$

1 Answer 1

3
$\begingroup$

Since this is model-based RL you can use policy iteration method of dynamic programming, and since this MDP is very simple you can intuitively guess an optimal policy for your initial iterated policy $\pi_0$ as $\pi_0(a_1|s_1)=1, \pi_0(a_2|s_2)=1$, where $a_1$ is the switch action and $a_2$ is the stay action.

In the first policy evaluation step we need to solve for the Bellman equation of both states:

$v_{\pi_0}(s_1)=\sum_a\pi_0(a|s_1)\big[\sum_rp(r|s_1,a_1)r+\gamma \sum_{s'}p(s'|s_1,a_1)v_{\pi_0}(s')\big]=0+0.5(0.2v_{\pi_0}(s_1)+0.8v_{\pi_0}(s_2))$

$v_{\pi_0}(s_2)=\sum_a\pi_0(a|s_2)\big[\sum_rp(r|s_2,a_2)r+\gamma \sum_{s'}p(s'|s_2,a_2)v_{\pi_0}(s')\big]=1+0.5v_{\pi_0}(s_2)$

This is a simple system of linear equations of 2 variables $v_{\pi_0}(s_1),v_{\pi_0}(s_2)$, you can easily check the solutions are $v_{\pi_0}(s_1)=\frac{8}{9},v_{\pi_0}(s_2)=2$.

Then in the first policy improvement step we need to compute action value for each state-action pair:

$q_{\pi_0}(s_1,a_1)=\sum_rp(r|s_1,a_1)r+\gamma \sum_{s'}p(s'|s_1,a_1)v_{\pi_0}(s')=0+0.5(0.2v_{\pi_0}(s_1)+0.8v_{\pi_0}(s_2))=\frac{8}{9}$

$q_{\pi_0}(s_1,a_2)=\sum_rp(r|s_1,a_2)r+\gamma \sum_{s'}p(s'|s_1,a_2)v_{\pi_0}(s')=0+0.5v_{\pi_0}(s_1)=\frac{4}{9}$

$q_{\pi_0}(s_2,a_1)=\sum_rp(r|s_2,a_1)r+\gamma \sum_{s'}p(s'|s_2,a_1)v_{\pi_0}(s')=0+0.5(0.2v_{\pi_0}(s_2)+0.8v_{\pi_0}(s_1))=\frac{5}{9}$

$q_{\pi_0}(s_2,a_2)=\sum_rp(r|s_2,a_2)r+\gamma \sum_{s'}p(s'|s_2,a_2)v_{\pi_0}(s')=1+0.5v_{\pi_0}(s_2)=2$

Therefore the improved policy is updated as $\pi_1(a_1|s_1)=1,\pi_1(a_2|s_2)=1$, which apparently already converges to $\pi_0$ initialized in the above first section. So we can conclude an optimal policy for this MDP is $\pi^*(a_1|s_1)=1, \pi^*(a_2|s_2)=1$.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .