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.

Python 3.5, PyTorch >= 0.2.0, numpy, gym, universe, cv2.

• 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.

• 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.