rl-sandbox

Personal sandbox project for testing reinforcement learning algorithms.

Setup

Install dependencies

It is recommended to use a conda environment. To create a new environment, run:

conda create -n rl-sandbox python=3.8

Activate the environment:

conda activate rl-sandbox

Install dependencies:

pip install swig
pip install -r requirements.txt

Run the code

The code is organized by environment and method, for example lunar_lander/A2C contains the code for the A2C algorithm on the Lunar Lander environment. To run or train an algorithm, navigate to the corresponding directory and run:

python train.py --resume [INSTANCE_NAME] --save_instance [INSTANCE_NAME]

Where INSTANCE_NAME is the name of the instance to run or resume. If INSTANCE_NAME is not provided, a new instance will be created. You can omit the --resume flag to start a new instance. --save_instance is used to save the instance to a file under env/method/instances, this argment is required.

You can tweak the parameters of the algorithm by editing the parameters.py file in the corresponding directory. Here you can toggle render_environment to render the environment while training, as well as toggle save_model to save the model after training

Plot instance results

To plot the results of an instance, run:

python plot.py --instance [INSTANCE_NAME]

TODO

• Update gym version to maintened fork gymnasium
• Refactor code to remove duplicate code
• Solve atari games