Proximal Policy Optimization - PPO

This is a PyTorch implementation of Proximal Policy Optimization - PPO.

PPO is a policy gradient method for reinforcement learning. Simple policy gradient methods do a single gradient update per sample (or a set of samples). Doing multiple gradient steps for a single sample causes problems because the policy deviates too much, producing a bad policy. PPO lets us do multiple gradient updates per sample by trying to keep the policy close to the policy that was used to sample data. It does so by clipping gradient flow if the updated policy is not close to the policy used to sample the data.

You can find an experiment that uses it here. The experiment uses Generalized Advantage Estimation.

26import torch
27
28from labml_helpers.module import Module
29from labml_nn.rl.ppo.gae import GAE

PPO Loss

Here’s how the PPO update rule is derived.

We want to maximize policy reward $\max_\theta J(\pi_\theta) = \mathop{\mathbb{E}}_{\tau \sim \pi_\theta}\Biggl[\sum_{t=0}^\infty \gamma^t r_t \Biggr]$ where $r$ is the reward, $\pi$ is the policy, $\tau$ is a trajectory sampled from policy, and $\gamma$ is the discount factor between $[0, 1]$.

$\begin{align} \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[ \sum_{t=0}^\infty \gamma^t A^{\pi_{OLD}}(s_t, a_t) \Biggr] &= \\ \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[ \sum_{t=0}^\infty \gamma^t \Bigl( Q^{\pi_{OLD}}(s_t, a_t) - V^{\pi_{OLD}}(s_t) \Bigr) \Biggr] &= \\ \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[ \sum_{t=0}^\infty \gamma^t \Bigl( r_t + V^{\pi_{OLD}}(s_{t+1}) - V^{\pi_{OLD}}(s_t) \Bigr) \Biggr] &= \\ \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[ \sum_{t=0}^\infty \gamma^t \Bigl( r_t \Bigr) \Biggr] - \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[V^{\pi_{OLD}}(s_0)\Biggr] &= J(\pi_\theta) - J(\pi_{\theta_{OLD}}) \end{align}$

So, $\max_\theta J(\pi_\theta) = \max_\theta \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[ \sum_{t=0}^\infty \gamma^t A^{\pi_{OLD}}(s_t, a_t) \Biggr]$

Define discounted-future state distribution, $d^\pi(s) = (1 - \gamma) \sum_{t=0}^\infty \gamma^t P(s_t = s | \pi)$

Then, $\begin{align} J(\pi_\theta) - J(\pi_{\theta_{OLD}}) &= \mathbb{E}_{\tau \sim \pi_\theta} \Biggl[ \sum_{t=0}^\infty \gamma^t A^{\pi_{OLD}}(s_t, a_t) \Biggr] \\ &= \frac{1}{1 - \gamma} \mathbb{E}_{s \sim d^{\pi_\theta}, a \sim \pi_\theta} \Bigl[ A^{\pi_{OLD}}(s, a) \Bigr] \end{align}$

Importance sampling $a$ from $\pi_{\theta_{OLD}}$,

$\begin{align} J(\pi_\theta) - J(\pi_{\theta_{OLD}}) &= \frac{1}{1 - \gamma} \mathbb{E}_{s \sim d^{\pi_\theta}, a \sim \pi_\theta} \Bigl[ A^{\pi_{OLD}}(s, a) \Bigr] \\ &= \frac{1}{1 - \gamma} \mathbb{E}_{s \sim d^{\pi_\theta}, a \sim \pi_{\theta_{OLD}}} \Biggl[ \frac{\pi_\theta(a|s)}{\pi_{\theta_{OLD}}(a|s)} A^{\pi_{OLD}}(s, a) \Biggr] \end{align}$

Then we assume $d^\pi_\theta(s)$ and $d^\pi_{\theta_{OLD}}(s)$ are similar. The error we introduce to $J(\pi_\theta) - J(\pi_{\theta_{OLD}})$ by this assumption is bound by the KL divergence between $\pi_\theta$ and $\pi_{\theta_{OLD}}$. Constrained Policy Optimization shows the proof of this. I haven’t read it.

$\begin{align} J(\pi_\theta) - J(\pi_{\theta_{OLD}}) &= \frac{1}{1 - \gamma} \mathop{\mathbb{E}}_{s \sim d^{\pi_\theta} \atop a \sim \pi_{\theta_{OLD}}} \Biggl[ \frac{\pi_\theta(a|s)}{\pi_{\theta_{OLD}}(a|s)} A^{\pi_{OLD}}(s, a) \Biggr] \\ &\approx \frac{1}{1 - \gamma} \mathop{\mathbb{E}}_{\color{orange}{s \sim d^{\pi_{\theta_{OLD}}}} \atop a \sim \pi_{\theta_{OLD}}} \Biggl[ \frac{\pi_\theta(a|s)}{\pi_{\theta_{OLD}}(a|s)} A^{\pi_{OLD}}(s, a) \Biggr] \\ &= \frac{1}{1 - \gamma} \mathcal{L}^{CPI} \end{align}$

32class ClippedPPOLoss(Module):

133    def __init__(self):
134        super().__init__()

136    def __call__(self, log_pi: torch.Tensor, sampled_log_pi: torch.Tensor,
137                 advantage: torch.Tensor, clip: float) -> torch.Tensor:

ratio $r_t(\theta) = \frac{\pi_\theta (a_t|s_t)}{\pi_{\theta_{OLD}} (a_t|s_t)}$; this is different from rewards $r_t$.

140        ratio = torch.exp(log_pi - sampled_log_pi)

Cliping the policy ratio

$\begin{align} \mathcal{L}^{CLIP}(\theta) = \mathbb{E}_{a_t, s_t \sim \pi_{\theta{OLD}}} \biggl[ min \Bigl(r_t(\theta) \bar{A_t}, clip \bigl( r_t(\theta), 1 - \epsilon, 1 + \epsilon \bigr) \bar{A_t} \Bigr) \biggr] \end{align}$

The ratio is clipped to be close to 1. We take the minimum so that the gradient will only pull $\pi_\theta$ towards $\pi_{\theta_{OLD}}$ if the ratio is not between $1 - \epsilon$ and $1 + \epsilon$. This keeps the KL divergence between $\pi_\theta$ and $\pi_{\theta_{OLD}}$ constrained. Large deviation can cause performance collapse; where the policy performance drops and doesn’t recover because we are sampling from a bad policy.

Using the normalized advantage $\bar{A_t} = \frac{\hat{A_t} - \mu(\hat{A_t})}{\sigma(\hat{A_t})}$ introduces a bias to the policy gradient estimator, but it reduces variance a lot.

169        clipped_ratio = ratio.clamp(min=1.0 - clip,
170                                    max=1.0 + clip)
171        policy_reward = torch.min(ratio * advantage,
172                                  clipped_ratio * advantage)
173
174        self.clip_fraction = (abs((ratio - 1.0)) > clip).to(torch.float).mean()
175
176        return -policy_reward.mean()

Clipped Value Function Loss

Similarly we clip the value function update also.

$\begin{align} V^{\pi_\theta}_{CLIP}(s_t) &= clip\Bigl(V^{\pi_\theta}(s_t) - \hat{V_t}, -\epsilon, +\epsilon\Bigr) \\ \mathcal{L}^{VF}(\theta) &= \frac{1}{2} \mathbb{E} \biggl[ max\Bigl(\bigl(V^{\pi_\theta}(s_t) - R_t\bigr)^2, \bigl(V^{\pi_\theta}_{CLIP}(s_t) - R_t\bigr)^2\Bigr) \biggr] \end{align}$

Clipping makes sure the value function $V_\theta$ doesn’t deviate significantly from $V_{\theta_{OLD}}$.

179class ClippedValueFunctionLoss(Module):

200    def __call__(self, value: torch.Tensor, sampled_value: torch.Tensor, sampled_return: torch.Tensor, clip: float):
201        clipped_value = sampled_value + (value - sampled_value).clamp(min=-clip, max=clip)
202        vf_loss = torch.max((value - sampled_return) ** 2, (clipped_value - sampled_return) ** 2)
203        return 0.5 * vf_loss.mean()