Unsupervised On-Policy Reinforcement Learning

This work combines Active Pre-Training with an On-Policy algorithm, Phasic Policy Gradient.

Active Pre-Training

Is used to pre-train a model free algorithm before defining a downstream task. It calculates the reward based on an estimatie of the particle based entropy of states. This reduces the training time if you want to define various tasks - i.e. robots for a warehouse.

Phasic Policy Gradient

Improved Version of Proximal Policy Optimization, which uses auxiliary epochs to train shared representations between the policy and a value network.