When deriving the Bellman equation for $q_\pi(s,a)$, we have

$q_\pi(s,a) = E_\pi[G_t | S_t = s, A_t = a] = E_\pi[R_{t+1} + \gamma G_{t+1} | S_t = s, A_t = a]$ (1)

This is what is confusing me, at this point, for the Bellman equation for $q_\pi(s,a)$, we write $G_{t+1}$ as an expected value, conditioned on $s'$ and $a'$ of the action value function at $s'$, otherwise, there is no recursion with respect to $q_\pi(s,a)$, and therefore no Bellman equation. Namely,

$ = \sum_{a \in A} \pi(a |s) \sum_{s' \in S} \sum_{r \in R} p(s',r|s,a)(r + \gamma E_\pi[G_{t+1}|S_{t+1} = s', A_{t+1} = a'])$ (2)

which introduces the recursion of $q$,

$ = \sum_{a \in A} \pi(a |s) \sum_{s' \in S} \sum_{r \in R} p(s',r|s,a)(r + \gamma q_\pi(s',a'))$ (3)

which should be the Bellman equation for $q_\pi(s,a)$, right?

On the other hand, when connecting $q_\pi(s,a)$ with $v_\pi(s')$, in this answer, I believe this is done

$q_\pi(s,a) = \sum_{a\in A} \pi(a |s) \sum_{s' \in S}\sum_{r \in R} p(s',r|s,a)(r + \gamma E_{\pi}[G_{t+1} | S_{t+1} = s'])$ (4)

$q_\pi(s,a) = \sum_{a\in A} \pi(a |s) \sum_{s' \in S}\sum_{r \in R} p(s',r|s,a)(r + \gamma v_\pi(s'))$ (5)

Is the difference between using the expectation $E_{\pi}[G_{t+1} | S_{t+1} = s', A_{t+1} = a']$ in (3) and the expectation $E_{\pi}[G_{t+1} | S_{t+1} = s']$ in $(4)$ simply the difference in how we choose to express the expected return $G_{t+1}$ at $s'$ in the definition of $q_\pi(s,a)$?

In $3$, we express the total return at $s'$ using the action value function

enter image description here

leading to the recursion and the Bellman equation, and in $4$, the total return is expressed at $s'$ using the value function

enter image description here

leading to $q_\pi(s,a) = q_\pi(s,a,v_\pi(s'))$?


1 Answer 1


Your understanding of the Bellman equation is not quite right. The state-action value function is defined as the expected (discounted) returns when taking action $a$ in state $s$. Now, in your equation (2) you have conditioned on taking action $a'$ in the inner expectation - this is not what happens in the state-action value function, you do not condition on knowing $A_{t+1}$, it is chosen according to the policy $\pi$ as per the definition of a bellman equation.

If you want to see a 'recursion' between state action value functions, note that

$$v_\pi(s) = \sum_a \pi(a|s)q_\pi(s,a)\;,$$

Your equation (5) is incorrect -- you need to drop the outter sum over $a$ as we have conditioned on knowing $a$. I will drop the $\pi$ subscripts for ease on notation, and we can see a 'recursion' for state-action value functions as:

$$q(s,a) = \sum_{s',r}p(s',r|s,a)\left(r + \gamma \left[\sum_{a'} \pi(a'|s')q(s',a')\right]\right)\;.$$


You must log in to answer this question.

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