Deep Reinforcement Learning with Online Generalized Advantage Estimation

October 6, 2016

Deep Reinforcement Learning, or Deep RL, is a really hot field at the moment. If you haven’t heard of it, pay attention. Combining the power of reinforcement learning and deep learning, it is being used to play complex games better than humans, control driverless cars, optimize robotic decisions and limb trajectories, and much more. And we haven’t even gotten started… Deep RL has far reaching applications in business, finance, health care, and many other fields which could be improved with better decision making. It’s the closest (practical) approach we have to AGI. Seriously… how cool it that? In this post, I’ll rush through the basics and terminology in standard reinforcement learning (RL) problems, then review and extend work in Policy Gradient and Actor-Critic methods to derive an online variant of Generalized Advantage Estimation (GAE) using eligibility traces, which can be used to learn optimal policies for our Deep RL agents.

Background and Terminology

The RL loop: state, action, reward

The RL framework: An agent senses the current state (s) of its environment. The agent takes an action (a), selected from the action set (A), using policy (π), and receives an immediate reward (r), with the goal of receiving large future return (R).

That’s it. Everything else is an implementation detail. The above paragraph can be summed up with the equations below, where $\sim$ means that we randomly sample a value from a probability distribution, and the current discrete time-step is $t$.

\[\begin{align} a_t &\sim \pi(a_t \mid s_t) \\ a_t &\in A_t \\ r_t, s_{t+1} &\sim p(r_t, s_{t+1} \mid s_t, a_t) \\ \tau_t &= \{ s_t, a_t, r_t, s_{t+1}, a_{t+1}, r_{t+1}, ~ ... \} \\ \tau &= \tau_0 \\ R(\tau_t) &= \sum_{l=0}^\infty{\gamma^l r_{t+l}} \end{align}\]

The policy is some function (usually parameterized, sometimes differentiable) which uses the full history of experience of the agent to choose an action after presentation of the new state of the environment. For simplicity, we’ll only consider discrete-time episodic environments, though most of this could be extended or approximated for continuous, infinite-horizon scenarios. We will also only consider the function approximation case, where we will approximate policies and value functions using neural networks that are parameterized by a vector of learnable weights $Θ$.

The difficulty of reinforcement learning, as compared to other sub-fields of machine learning, is the Credit Assignment Problem. This is the problem that there are many many actions which lead to any given reward (and many rewards resulting from a single action), and it’s not easy to pick out the “important actions” which led to the good (or bad) reward. With infinite resources and enough samples, the statistics will reveal the truth, but we’d like to make sure an algorithm converges in our lifetime.

Approaches to RL

The goal with all reinforcement learners is to learn a good policy which will take the best actions in order to generate the highest return. This can be accomplished a few different ways.

Value iteration, temporal difference (TD) learning, and Q-learning approximate the value of states $V(s_t)$ or state-action pairs $Q(s_t, a_t)$, and use the “surprise” from incorrect value estimates to update a policy. These methods can be more flexible for learning off-policy, but can be difficult to apply effectively for continuous action spaces (such as robotics).

An alternative approach, called Policy Gradient methods, model a direct mapping from states to actions. With a little math, one can compute the effect of a parameter $\theta_i$ on the future cumulative returns of the policy. Then to improve our policy, we “simply” adjust our parameters in a direction that will increase the return. In truth, what we’re doing is increasing the probability of choosing good actions, and decreasing the probability of choosing bad actions.

I put “simply” in quotes because, although the math is reasonably straightforward, there are a few practical hurdles to overcome to be able to learn policies effectively. Policy gradient methods can be applied to a wide range of problems, with both continuous and discrete action spaces. In the next section we’ll see a convenient theorem that makes the math of computing parameter gradients tractable.

We can combine value iteration and policy gradients using a framework called Actor-Critic. In this, we maintain an actor which typically uses a policy gradient method, and a critic, which typically uses some sort of value iteration. The actor never sees the actual reward. Instead, the critic intercepts the rewards from the trajectory and critiques the chosen actions of the actor. This way, the noisy (and delayed) reward stream can be smoothed and summarized for better policy updates. Below, we will assume an Actor-Critic approach, but we will focus only on how to update the parameters of the actor.

Policy Gradient Theorem

I’d like to briefly review a “trick” which makes policy gradient methods tractable: the Policy Gradient Theorem. If we assume there is some mapping of any state/action pair to a real-valued return:

\[f: S \times A \rightarrow \mathbb{R} \\ x \in (S,A) \\ R = f(x)\]

then we’d like to compute the gradient of $\theta$ with respect to the expected total return $E_x[f(x)]$:

\[\begin{align} \nabla_\theta E_x[f(x)] &= \frac{\partial E_x[f(x)]}{\partial \theta} \\ & = \nabla_\theta \int f(x) ~ p(x \mid \theta) ~ dx \\ & = \int f(x) ~ \nabla_\theta p(x \mid \theta) ~ \frac{p(x \mid \theta)}{p(x \mid \theta)} ~ dx \\ & = \int f(x) ~ \nabla_\theta log ~ p(x \mid \theta) ~ p(x \mid \theta) ~ dx \\ & = E_x[f(x) ~ \nabla_\theta log ~ p(x \mid \theta)] \end{align}\]

We bring the gradient inside the integral and divide by the probability to get the gradient of log probability (grad-log-prob) term, along with a well-formed expectation. This trick means we don’t actually need to compute the integral equation in order to estimate the total gradient. Additionally, it’s much easier to take the gradient of a log as it decomposes into a sum of terms. This is important, because we only need to sample rewards and compute $\nabla_\theta log ~ p(x \mid \theta)$, which is dependent solely on our policy and the states/rewards that we see. There’s no need to understand or estimate the transition probability distribution of the underlying environment! (this is called “model-free reinforcement learning”.) Check out John Schulman’s lectures for a great explanation and more detail.

Improvements to the policy gradient formula

Now we have this easy-to-calculate formula to estimate the gradient, so we can just plug in the formulas and feed it into a gradient descent optimization, and we’ll magically learn optimal parameters… right? Right?!?

Sadly, due to weak correlation of actions to rewards resulting from the credit assignment problem, our gradient estimates will be extremely weak and noisy (the signal to noise ratio will be small, and variance of the gradient will be high), and convergence will take a while (unless it diverges).

One improvement we can make is to realize that, when deciding which actions to select, we only care about the relative difference in state-action values. Suppose we’re in a really awesome state, and no matter what we do we’re destined to get a reward of at least 100, but there’s a single action which would allow us to get 101. In this case we care only about that difference: \(101 - 100 = 1\).

This is the intuition behind the advantage function $A^\pi(s,a)$ for a policy $\pi$. It is defined as the difference between the state-action value function $Q^\pi(s,a)$ and the state value function $V^\pi(s)$:

\[\begin{align} V^\pi(s_t) &= E^\pi[R(\tau_t) \mid s_t] \\ Q^\pi(s_t, a_t) &= E^\pi[R(\tau_t) \mid s_t, a_t] \\ A^\pi(s_t, a_t) &= Q^\pi(s_t, a_t) - V^\pi(s_t) \end{align}\]

Going back to the policy gradient formula, we can subtract a zero-expectation baseline $b_t(s_t)$ from our score function $f(x)$ without changing the expectation. If we choose the score function to be the state-action value function $Q^\pi(s,a)$, and the baseline to be the state value function $V^\pi(s)$, then the policy gradient $g$ is of the form:

\[\begin{align} g = E[\sum_{t=0}^\infty{A^\pi(s_t, a_t) ~ \nabla_\theta log ~ \pi^\theta(a_t \mid s_t)}] \end{align}\]

The intuition with this formula is that we wish to increase the probability of better-than-average actions, and decrease the probability of worse-than-average actions. We use a discount factor $\gamma$ to control the impact of our value estimation. With $\gamma$ near 0, we will approximate the next-step return. With $\gamma$ near 1, we will approximate the sum of all future rewards. If episodes are very long (or infinite), we may need $\gamma < 1$ for tractibility/convergence.

This formula is an example of an Actor-Critic algorithm, where the policy $\pi$ (the actor) adjusts its parameters by using the “advice” of a critic. In this case we use an estimate of the advantage to critique our policy choices.

Generalized Advantage Estimator (GAE)

Schulman et al use a discounted sum of TD residuals:

\[\begin{align} \delta_t^V = r_t + \gamma V(s_{t+1}) - V(s_t) \end{align}\]

and compute an estimator of the k-step discounted advantage:

\[\begin{align} \hat{A}_t^{(k)} = \sum_{l=0}^{k-1}{\gamma^l \delta_{t+l}^V} \end{align}\]

Note: it seems that equation 14 from their paper has an incorrect subscript on $\delta$.

They define their generalized advantage estimator (GAE) as the weighted average of the advantage estimators above, which reduce to a sum of discounted TD residuals:

\[\begin{align} \hat{A}_t^{GAE(\gamma,\lambda)} = \sum_{l=0}^\infty{(\gamma\lambda)^l \delta_{t+l}^V} \end{align}\]

This generalized estimator of the advantage function allows a trade-off of bias vs variance using the parameter $0 \leq \lambda \leq 1$, similar to TD(λ). For $\lambda = 0$, the problem reduces to the (unbiased) TD(0) function. As we increase $\lambda$ towards 1, we reduce the variance of our estimator but increase the bias.

Online GAE

We’ve come a long way. We now have a low(er)-variance approximation to the true policy gradient. In github speak: :tada: But for problems I care about (for example high frequency trading strategies), it’s not very practical to compute forward-looking infinite-horizon returns before performing a gradient update step.

In this section, I’ll derive an online formula which is equivalent to the GAE policy gradient above, but which uses eligibility traces of the inner gradient of log probabilities to compute a gradient estimation on every reward, as it arrives. Not only will this prove to be more efficient, but there will be a massive savings in memory requirements and compute resources, at the cost of a slightly more complex learning process. (Luckily, I code in Julia)

Notes: See Wawrzyński et al for an alternate derivation. These methods using eligibility traces are closely related to REINFORCE.

Lets get started. We are trying to reorganize the many terms of the policy gradient formula so that the gradient is of the form: $g = E^\pi[\sum_{t=0}^\infty{r_t \psi_t}]$, where $\psi_t$ can depend only on the states, actions, and rewards that occurred before (or immediately after) the arrival of $r_t$. We will solve for an online estimator of the policy gradient $\hat{g}$:

\[\begin{align} \hat{g} &= \sum_{t=0}^\infty{\hat{A}_t^{GAE(\gamma,\lambda)} ~ \nabla_\theta log ~ \pi^\theta(a_t \mid s_t)} \\ & = \sum_{t=0}^\infty{ \nabla_\theta log ~ \pi^\theta(a_t \mid s_t) ~ \sum_{l=0}^\infty{(\gamma\lambda)^l \delta_{t+l}^V}} \end{align}\]

In order to simplify the derivation, I’ll introduce the following shorthand:

\[\begin{align} \nabla_t := \nabla_\theta log ~ \pi^\theta(a_t \mid s_t) \end{align}\]

and then expand the sum:

\[\begin{align} \hat{g} &= \nabla_0 (\delta_0^V + (\gamma\lambda)\delta_1^V + (\gamma\lambda)^2\delta_2^V + ~ ... ) \nonumber \\ & ~~ + \nabla_1 (\delta_1^V + (\gamma\lambda)\delta_2^V + (\gamma\lambda)^2\delta_3^V + ~ ... ) \\ & ~~ + \nabla_2 (\delta_2^V + (\gamma\lambda)\delta_3^V + (\gamma\lambda)^2\delta_4^V + ~ ... ) \nonumber \\ & ~~ + ~ ... \nonumber \end{align}\]

and collect the $\delta_t^V$ terms:

\[\begin{align} \hat{g} &= \delta_0^V \nabla_0 \nonumber \\ & ~~ + \delta_1^V ( \nabla_1 + (\gamma\lambda)\nabla_0 ) \\ & ~~ + \delta_2^V ( \nabla_2 + (\gamma\lambda)\nabla_1 + (\gamma\lambda)^2\nabla_0 ) \nonumber \\ & ~~ + ~ ... \nonumber \end{align}\]

and summarize:

\[\begin{align} \hat{g} &= \sum_{t=0}^\infty{\delta_t^V \sum_{l=0}^t{(\gamma\lambda)^l \nabla_{t-l}}} \end{align}\]

If we define our eligibility trace as the inner sum in that equation:

\[\begin{align} \epsilon_t := \sum_{l=0}^t{(\gamma\lambda)^l \nabla_{t-l}} \end{align}\]

and convert to a recursive formula:

\[\begin{align} \epsilon_0 &:= \nabla_0 \\ \epsilon_t &:= (\gamma\lambda) \epsilon_{t-1} + \nabla_t \end{align}\]

then we have our online generalized advantage estimator for the policy gradient:

\[\begin{align} \hat{g} = \sum_{t=0}^\infty{\delta_t^V \epsilon_t} \end{align}\]

So at each time-step, we compute the gradient term $\hat{g}_t = \delta_t^V \epsilon_t$ as the product of the TD(0) error from our critic and the accumulated log-prob gradients of our policy. We could update our parameters online at the end of each episode, or at each step, or in batches, or using some sort of smoothing method.


In this post I covered some basic terminology, and summarized the theorems and formulas that comprise Generalized Advantage Estimation. We extended GAE to the online setting in order to give us more flexibility when computing parameter updates. The hyperparameter $\gamma$ allows us to control our trust in the value estimation, while the hyperparameter $\lambda$ allows us to assign more credit to recent actions.

In future posts, I intend to demonstrate how to use formulas 26-28 in practical algorithms to solve problems in robotic simulation and other complex environments. If you need help solving difficult problems with data, or if you see mutual benefit in collaboration, please don’t hesitate to get in touch.