PIXELBANKv8.2.1
Menu
Back to RL Study Plan
Week 8

Chapter 8: Actor-Critic & PPO

Combine value and policy learning: the actor-critic architecture, generalized advantage estimation, synchronous and asynchronous A2C/A3C, the trust-region idea behind TRPO, and the clipped surrogate objective that made PPO the default algorithm.

Chapter Overview

Chapter 7 ended with a working but impractical method: REINFORCE with a baseline is unbiased and general, yet needs complete episodes and still has high variance. Chapter 4 already supplied the fix — bootstrap. Replacing the sampled return with a learned value estimate gives the actor-critic architecture, where a policy (the actor) is improved using advantages supplied by a value function (the critic).

That combination is powerful but delicate. Policy gradient steps are only valid locally: the gradient is computed under the current policy's trajectory distribution, so a step large enough to change that distribution invalidates the very data it was computed from. TRPO addressed this with a hard constraint on how far the policy may move, at considerable implementation cost. PPO achieved nearly the same effect by clipping a probability ratio — a change of a few lines that turned out to be robust enough to become the default algorithm across robotics, games, and RLHF for language models.

Chapter Roadmap

Click any topic to jump in

1
Actor-Critic

A policy improved by advantages from a learned value function — bootstrapping brings policy gradients online.

Two independent problems remain
2
GAE

Exponentially weight n-step advantage estimates to tune the bias-variance tradeoff with a single parameter.

Constraining the step is the harder one

A hard KL constraint works but demands second-order optimization.

3
A2C / A3C

Run many environments at once so a batch of decorrelated transitions replaces the replay buffer.

4
Trust Regions & TRPO

Constrain the KL divergence between successive policies so each update stays in the region where the estimate is valid.

A first-order approximation suffices
5
PPO

Approximate the trust region by clipping the probability ratio — simple, robust, and the field's default.

Sign up to unlock this chapter

This chapter is part of PixelBank Premium. Create a free account, then upgrade to read the full lesson — concepts, walkthroughs, and exercises.