Implementing vanilla DQN model

A simple implementation in Pytorch of vanilla DQN model, featured in the paper Playing Atari with Deep Reinforcement Learning (2013) arXiv:1312.5602

Environment specifications:

Python version: 3.9.15
Default replay memory might takes up a lot of memory, ~20 GBs
Dependencies:

Package	Version	Installation note
gym	0.21.0	N/A
torch	2.1.0.dev20230526	refer to official installation page
tensorboard	2.11.0	N/A
matplotlib	3.6.2	N/A

** No installation note means package can simply be installed via pip

---

Demo

Pong

Model after 2,500,000 training steps (Learning rate=0.00025)

pong_demo.mp4

Breakout

Model after 24,000,000 training steps (Learning rate=0.000025, wihout penalizing lose lives)
Note: The model prone to perform random actions instead of FIRE (to reset game) after losing lives (which was cut out by force reset)

breakout_nopen_demo.mp4

Model (Learning rate=0.000025, with penalizing lose lives) (to be updated)
N/A