luig-io is a collection of deep reinforcement learning algorithms implemented using the Keras library. My implementations focus on readability, following Keras best practices, and producing idiomatic abstractions.
The algorithms are geared at solving the gym-super-mario-bros environment.
This section shows how to begin working with the codebase.
First, you will need to install the luig_io library:
python setup.py developTo verify that the installation worked, lets run the policy_gradient algorithm:
python algorithms/reinforce/main.pyIf your environment is properly configured, you will begin to see log statements.
More information about each algorithm and it's corresponding setup can be found each directories' corresponding README.
The project is structured into a core library, luig_io, and individual algorithms.
The luig_io package contains helpers and share utilities.
These include simple models, such as a simple 3 layer: SimpleCNN, gym wrappers to implement FrameStack, GrayScale, and Resize operations.
These are used across all environments.
The luig_io package also contains the agents and their components.
The entrypoints contain code to run the actual models. This includes helpers, environment loading and wrapping to prepare for the algorithm, training code, sample collection code, and anything else needed to run an algorithm.
The repo contains the following algorithms
- REINFORCE (aka vanilla Policy Gradient)
- ❌ PPO (in progress)
- ❌ WorldModels
- ❌ Deep Q Learning
- ... and more!
Contributions are welcome. Send them over via a PR.
If any code in this repository is useful in any of your research, please consider citing the codebase:
@misc{wood2022luigio,
title={Luig-IO},
author={Wood, Luke},
year={2022},
howpublished={\url{https://github.com/lukewood/luig-io}},
}