Skip to content

Latest commit

 

History

History
9 lines (7 loc) · 854 Bytes

File metadata and controls

9 lines (7 loc) · 854 Bytes

Advantage-Actor-Critic for OpenAI-gym environments

Implementation of Advantage-Actor-Critic with entropy regularization in Pytorch for OpenAI-gym environments.

Advantage-Actor-Critic

The policy gradient in Adavantage-Actor-Crititc differes from the classical REINFORCE policy gradient by using a baseline to reduce variance. This baseline is an approximation of the state value function (Critic). Since the baseline is not dependent on the action this does not introduce bias.
For more detailed information I would recommend reading this articel.

Entropy regularization

In order to encourage exploration we add the entropy of the policy distribution to the loss. This forces the actor to consider as much actions as possible while still maximizing the reward.