In the policy gradient method, there's a trick to reduce the variance of policy gradient. We use causality, and remove part of the sum over rewards so that only actions happened after the reward are taken into account (See here http://rail.eecs.berkeley.edu/deeprlcourse/static/slides/lec-5.pdf, slide 18).

Why does it work? I understand the intuitive explanation, but what's the rigorous proof of it? Can you point me to some papers?


1 Answer 1


An important thing we're going to need is what is called the "Expected Grad-Log-Prob Lemma here" (proof included on that page), which says that (for any $t$):

$$\mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta}(a_t \mid s_t) \right] = 0.$$

Taking the analytical expression of the gradient (from, for example, slide 9) as a starting point:

$$\begin{aligned} \nabla_{\theta} J(\theta) &= \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \left( \sum_{t=1}^T \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \right) \left( \sum_{t=1}^T r(s_t, a_t) \right) \right] \\ % &= \sum_{t=1}^T \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=1}^T r(s_{t'}, a_{t'}) \right] \\ % &= \sum_{t=1}^T \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=1}^{t-1} r(s_{t'}, a_{t'}) + \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=t}^T r(s_{t'}, a_{t'}) \right] \\ % &= \sum_{t=1}^T \left( \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=1}^{t-1} r(s_{t'}, a_{t'}) \right] \\ + \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=t}^T r(s_{t'}, a_{t'}) \right] \right) \\ \end{aligned}$$

At the $t^{th}$ "iteration" of the outer sum, the random variables $ \sum_{t'=1}^{t-1} r(s_{t'}, a_{t'}) $ and $ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) $ are independent (we assume, by definition, the action only depends on the most recent state), which means we are allowed to split the expectation:

$$\nabla_{\theta} J(\theta) = \sum_{t=1}^T \left( \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \sum_{t'=1}^{t-1} r(s_{t'}, a_{t'}) \right] \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \right] \\ + \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=t}^T r(s_{t'}, a_{t'}) \right] \right)$$

The first expectation can now be replaced by $0$ due to the lemma mentioned at the top of the post:

$$ \begin{aligned} \nabla_{\theta} J(\theta) % &= \sum_{t=1}^T \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \sum_{t'=t}^T r(s_{t'}, a_{t'}) \right] \\ % &= \mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \sum_{t=1}^T \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \left( \sum_{t'=t}^T r(s_{t'}, a_{t'}) \right). \\ \end{aligned} $$

The expression on slide 18 of the linked slides is an unbiased, sample-based estimator of this gradient:

$$\nabla_{\theta} J(\theta) \approx \frac{1}{N} \sum_{i=1}^N \sum_{t=1}^T \nabla_{\theta} \log \pi_{\theta} (a_{i, t} \mid s_{i, t}) \left( \sum_{t'=t}^T r(s_{i, t'}, a_{i, t'}) \right)$$

For a more formal treatment of the claim that we can pull $\sum_{t'=1}^{t-1} r(s_{t'}, a_{t'})$ out of an expectation due to the Markov property, see this page: https://spinningup.openai.com/en/latest/spinningup/extra_pg_proof1.html

  • $\begingroup$ Why is the form of "Expected Grad-Log-Prob Lemma" mentioned at the top of the post legit? It is different than the form of the proof shown in the link (the proof in the link has no problem). Your lemma is an expectation over trajectories. $\endgroup$
    – starriet
    Jan 4 at 3:09
  • $\begingroup$ @starriet For the individual choice of a single action at one point in time within a trajectory, the trajectory as a whole has no influence (our policy does not look at the history or the future, just at the current state), so $\mathbb{E}_{\tau \sim \pi_{\theta}(\tau)} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \right] = \mathbb{E}_{a_t \sim \pi_{\theta}} \left[ \nabla_{\theta} \log \pi_{\theta} (a_t \mid s_t) \right]$, where the right-hand side more closely matches that notation from the link. $\endgroup$
    – Dennis Soemers
    Jan 4 at 9:43
  • $\begingroup$ @starriet However, we use the trajectory-based notation because in the larger equation, we have these sums over all time steps, and that's the part where the full trajectory actually becomes relevant. $\endgroup$
    – Dennis Soemers
    Jan 4 at 9:45
  • $\begingroup$ Thanks, I think I got your point. If I understood correctly, the right-hand side is at a specific state, and the left-hand side is at the various states and actions of all possible trajectories, right? Since the RHS is 0 and it's about one state, LHS is also 0 because LHS is an expectation over all states (this is how I understood. please let me know if I'm wrong). $\endgroup$
    – starriet
    Jan 4 at 12:41
  • $\begingroup$ But still, It's difficult to understand why the form of the LHS is legit. It's an expectation over trajectories, but the term inside of the square bracket is a function of action and state. Given that a trajectory is composed of many states and actions, how the expectation over trajectories is possible if the term inside of the brackets is smaller than a trajectory? $\endgroup$
    – starriet
    Jan 4 at 12:51

You must log in to answer this question.

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