patefon / dqn_yie-ar-kung-fu

Let's train an agent to play Yie Ar Kung-Fu using DQN

AI Kung-Fu

---

This is a demo project built with a goal to explore in more details computer vision (CV) and reincorcement learning (RL) and has no any commercial use.

agent gameplay / youtube

Requirements:

• Linux Debian / Ubuntu / Mint
• python 3.8 / pytorch / openCV
• nestopia (as a NES emulator)

Goals

• Ultimate goal: beat 50+ levels.
• Current result: no more than 5 levels.

An approach:

• Q-learning / Deep Q Network
• do not use frameworks like OpenAI Gym, just plain python and torch as dep.

Run an app:

# 1) install Nestopia (NES emulator):
~ sudo apt-get update && apt-get install Nestopia
# 2) Place a rom with a game inside [./roms] dir.
# 3) Init python venv and install deps:
~ python3 -m venv [path_to_venv_dir]
~ source [path_to_venv_dir]/bin/activate
~ pip install -r requirements.txt
# 4) Check [./consts.py] and make changes if it necessary
# 5) Run an app
~ (venv) python app.py

Overview of dependencies:

Name	Version	License
Cython	0.29.23	Apache Software License
cycler	0.10.0	BSD
dataclasses	0.6	Apache Software License
evdev	1.4.0	BSD License
future	0.18.2	MIT License
imutils	0.5.3	MIT License
kiwisolver	1.3.1	BSD License
matplotlib	3.4.1	Python Software Foundation License
mss	6.1.0	MIT License
numpy	1.19.4	BSD
opencv-python	4.5.1.48	MIT License
Pillow	8.2.0	Historical Permission Notice and Disclaimer (HPND)
pyparsing	2.4.7	MIT License
pynput	1.7.3	LGPLv3
python-dateutil	2.8.1	BSD License
six	1.15.0	MIT License
torch	1.7.0	BSD License
torchvision	0.8.0	BSD
typing-extensions	3.7.4.3	Python Software Foundation License

References and thanks to:

• "Deep Reinforcment Learning Hands-on" book, Max Lapan