Proximal Policy Optimization

Basic Concept

$$\text{Critic Net} : \underbrace{\mathbb R^4}_\text{state vector} \to \underbrace{\mathbb R}_\text{state value}$$ $$\text{Actor Net} : \underbrace{\mathbb R^4}_\text{state vector} \to \underbrace{\mathbb R^2}_\text{action logits}$$

• Suppose that our environment has a 4-dimensional continuous state space and two discrete actions.

$$\text{Actor Net}(s_t; \theta) = \pi_\theta(a_t | s_t)$$

• So the actor network takes 4D state vectors and returns the logits of an action given a state.
• In other words, the actor network returns the distribution of an action given a state — which is the same as a policy function's role.

$$\text{Critic Net}(s_t; \phi) = V_\phi(s_t)$$

• And the critic network returns the value of state.

How to Train Actor Net (Policy)

• To train our actor network, first of all, we should define the loss function.
• It is define as follow :

$$\begin{align} L(s,a,\theta_k,\theta) &= -\min\left( \frac{\pi_{\theta}(a|s)}{\pi_{\theta_k}(a|s)} A^{\pi_{\theta_k}}(s,a), \text{clip}\left(\frac{\pi_{\theta}(a|s)}{\pi_{\theta_k}(a|s)}, 1 - \epsilon, 1+\epsilon \right) A^{\pi_{\theta_k}}(s,a)\right)\\\ &= -\min\left(\frac{\pi_{\theta}(a|s)}{\pi_{\theta_k}(a|s)} A^{\pi_{\theta_k}}(s,a),~g(\epsilon,~ A^{\pi_{\theta_k}}(s,a))\right) \end{align}$$

• where $\pi_\theta$ is our new policy, and $\pi_{\theta_k}$ is the old policy.

$$g(\epsilon, A) = \begin{cases} (1 + \epsilon) A & A \geq 0 \\\ (1 - \epsilon) A & A < 0 \end{cases}$$

• So we can decompose our loss function as follow :

$$-L(s,a,\theta_k,\theta) = \begin{cases} \min\left[ \frac{\pi_{\theta}(a|s)}{\pi_{\theta_k}(a|s)}, (1 + \epsilon) \right] \times A^{\pi_{\theta_k}}(s,a) & _{\text{if}~~A^{\pi_{\theta_k}} \ge 0} \\\ \max\left[ \frac{\pi_{\theta}(a|s)}{\pi_{\theta_k}(a|s)}, (1 - \epsilon) \right] \times A^{\pi_{\theta_k}}(s,a) & _{\text{if}~~A^{\pi_{\theta_k}} < 0} \end{cases}$$

• In this loss function equation, there's a kind of regularization for new policy ($\pi_\theta$).
• When the advantage is positive, the objective will increase if the action becomes more likely—that is, if $\pi_{\theta}(a|s)$ increases.
• But the min in this term puts a limit to how much the objective can increase.
• Once $\pi_{\theta}(a|s) &gt; (1+\epsilon) \pi_{\theta_k}(a|s) $, the min kicks in and this term hits a ceiling of $(1+\epsilon) A^{\pi_{\theta_k}}(s,a)$.
• Thus, the new policy does not benefit by going far away from the old policy.
• Likewise, when the advantage is negative, the objective will increase if the action becomes less likely—that is, if $\pi_{\theta}(a|s)$ decreases.
• But the max in this term puts a limit to how much the objective can increase.
• Once $\pi_{\theta}(a|s) &lt; (1-\epsilon) \pi_{\theta_k}(a|s)$, the max kicks in and this term hits a ceiling of $(1-\epsilon) A^{\pi_{\theta_k}}(s,a)$.
• Thus, again the new policy does not benefit by going far away from the old policy.

Advantage Estimates

• In our example code, our advantage estimate is computed as follow:

$$\delta_t = \underbrace{r_t + \gamma V_\phi(s_{t+1})}_{\text{the estimated value of}~ s_t} - \underbrace{V_\phi(s_t)}_\text{value func as baseline} \\$$$$ $$\hat A_t = \sum_{i=t}^T (\gamma \lambda )^{T-i}\delta_i$$

• where $r_t + \gamma V_\phi(s_{t+1})$ is the estimated state $s_t$ value which is based on the Bellman equation.

Rewards-to-go Etstimates

• Our reward-to-go estimate is computed as follow:

$$\hat R_t = \sum_{i=t}^T \gamma^{T-i} \underbrace{r_i}_\text{reward}$$

• where the reward $r_i$ is collected by our agent's journey.
• Basically our critic network, or value function, targets to predict a state value.
• And a state value is basically an expectation of a return, and the return is a discounted sum of future rewards.
• So, it makes sense to define the critic-net's loss as $\text{MSE}(V_\phi(s_t),~ \hat R_t)$

References

1. OpenAI Spinning Up : Proximal Policy Optimization.
2. Illias Chrysovergis (2021). "Proximal Policy Optimization " (Keras Tutorial).
3. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov (OpenAI) (2017). "Proximal Policy Optimization Algorithms".