Deep Reinforcement Learning Hands-On

Proximal Policy Optimization

Historically, this method came from the OpenAI team and was proposed long after TRPO (which is from 2015), but PPO is much simpler than TRPO, so we'll start from it. The paper in which it was proposed is by John Schulman et al and called Proximal Policy Optimization Algorithms, published in 2017 (arXiv:1707.06347).

The core improvement over the classical Asynchronous Advantage Actor-Critic (A3C) method is to change the expression used to estimate the PG. Instead of the gradient of logarithm probability of the action taken, the PPO method uses a different objective: the ratio between the new and the old policy scaled by the advantages.

In math form, the old A3C objective could be written as . The new objective proposed ...