https://github.com/tnfru/unsupervised-on-policy
Self supervised reward function + Phasic Policy Gradient
https://github.com/tnfru/unsupervised-on-policy
deep-reinforcement-learning reinforcement-learning self-supervised-learning unsupervised-learning unsupervised-reinforcement-learning
Last synced: about 1 month ago
JSON representation
Self supervised reward function + Phasic Policy Gradient
- Host: GitHub
- URL: https://github.com/tnfru/unsupervised-on-policy
- Owner: tnfru
- License: mit
- Created: 2021-07-11T22:07:50.000Z (almost 5 years ago)
- Default Branch: master
- Last Pushed: 2022-03-14T09:56:01.000Z (over 4 years ago)
- Last Synced: 2026-03-19T02:56:12.560Z (4 months ago)
- Topics: deep-reinforcement-learning, reinforcement-learning, self-supervised-learning, unsupervised-learning, unsupervised-reinforcement-learning
- Language: Python
- Homepage:
- Size: 348 KB
- Stars: 1
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Unsupervised On-Policy Reinforcement Learning
This work combines [Active Pre-Training](https://arxiv.org/abs/2103.04551) with an On-Policy algorithm, [Phasic Policy Gradient](https://arxiv.org/abs/2009.04416).
## 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.
## Installation
Install via pip `pip install unsupervised-on-policy`.
Configure by changing values in the config at `run_train.py`.
## Run
Supports the key words `load` `skip_pretrain` and `model_path`.
Example usage:
`python3 -m unsupervised_on_policy.run_train --skip_pretrain --prefix test
--model_path /home/lars/apt_ppg`