This project provides a clasical deep-q-learning strategy for solving diverse games. It uses gym-retro for the emulation of games and the environment. The files training.py and training-smb.py provides a framework for Super Mario World and Super Mario Bros games.
-
Supported Pythons:
- 3.5
- 3.6
-
Required packages:
- numpy, matplotlib, tensorflow, keras, docopt, gym-retro
These packages can be installed using pip (with
pip install <package_name>). However, tensorflow is no more available on pip. Usepip install --upgrade https://storage.googleapis.com/tensorflow/mac/cpu/tensorflow-1.12.0-py3-none-any.whl(see more here); be aware of your OS version to choose the right tensorflow's version. -
Install Super Mario World and Super Mario Bros The files in the directories SuperMarioX in this project (data and scenario json files) have to be put in the corresponding gym-retro directory.
- Where is located my gym-retro installation ?
Simply run a python interpreter and execute
import retro retro.__path__
to know your gym-retro installation location
- Where to put the scenarios and data files ?
Go to the directory found on the previous step (e.g.,
PATH_TO_PYTHON/python3.X/site-packages/retro) and and then go to the data/stable directory. All games supported by gym-retro are located here. In our case, go to SuperMarioWorld-Snes or SuperMarioBros-Nes and put the scenario et data json files in it. -
Run the training ?
Simply run training.py for SuperMarioWorld or training-smb.py for SuperMarioBros
-
Change the reward function
- The file data.json contains RAM addresses linked to some game informations.
- The file scenario.json contains the rewards of states reaching these game informations.