https://github.com/alro10/rl_pong
This is a repository for project in MO810 course-1s2018 IC-UNICAMP. The project is about implement DQN, ES and policy gradients for Pong and Catch game.
https://github.com/alro10/rl_pong
catch deep-reinforcement-learning dqn google-colab google-colaboratory pong-game
Last synced: 7 months ago
JSON representation
This is a repository for project in MO810 course-1s2018 IC-UNICAMP. The project is about implement DQN, ES and policy gradients for Pong and Catch game.
- Host: GitHub
- URL: https://github.com/alro10/rl_pong
- Owner: Alro10
- License: mit
- Created: 2018-06-12T18:55:22.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2018-06-19T20:44:09.000Z (over 7 years ago)
- Last Synced: 2025-01-20T13:27:28.287Z (9 months ago)
- Topics: catch, deep-reinforcement-learning, dqn, google-colab, google-colaboratory, pong-game
- Language: Jupyter Notebook
- Size: 183 KB
- Stars: 0
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# DQN-NES-Pong
This is a repository for project in MO810 course-1s2018 IC-UNICAMP. The project is about implement DQN, NES and policy gradients for Pong and Catch game.
## Requirements
Python 3.5, PyTorch >= 0.2.0, numpy, gym, universe, cv2.
## Deep Q-learning (DQN):
* `dqn_pong.ipynb` : This is a DQN implementation for Pong game (gym environment) and was trained in google colab (aprox 5 hours). Achieved **reward = 18**.
* `kerasdqn_catch.ipynb`: For learn more about DQN, I decided to implement a shallow neural network for catch enviroment. See the results in file.
## Evolution Strategies (Natural):
* `main.py` : Train ES on Pong and achieved **reward = 5** after 72 hours of training. Functions from `train.py`, `envs.py` , `model.py`
```
python3 main.py --env-name Pong-v4 --n 10 --lr 0.01 --useAdam
```
* `catch_ES.ipynb`: NES implementation for catch game! See results in flie.