Skip to content
Reinforcement Learning

Advanced policy optimization

Learning objective

Section titled “Learning objective”

Make policy gradient stable enough to trust: constrained updates and modern actor-critic.

Planned scope

Section titled “Planned scope”
  • A2C and A3C
  • Why unconstrained policy updates collapse
  • Trust regions and TRPO
  • PPO and the clipped objective
  • Generalised advantage estimation

Prerequisites

Section titled “Prerequisites”

Everything above this chapter in the sidebar.