topological-labyrinths-rl

Reinforcement Learning on Topological Labyrinth Environments

Setup

Set up python environment for this project (anaconda, venv, etc.) and install requirements:

pip install -r requirements.txt

please let me know if requirements are missing.

Usage

Running the training scripts

For now, we can train an agent with an MLP-Policy using the A2C, PPO and DQN algorithms. The implementations are provided by the stable-baselines3 library.

But first, we need to generate some mazes (it also comes with a help page, so just pass the -h flag to see how to customize):

python maze_sampler.py

It will generate a maze dictionary file at mazes/mazes_<lx>x<lz>_<maze_dict_key>.p. You can visualize the maze dictionary by executing

python viz_maze.py <path/to/mazes.p>

To run training with multiprocessing execute

mpiexec -n 5 python mpi_train.py <path/to/mazes.p>

This will train the agent and drop logs at results/experiments_<experiment_key>.

Visualizing results

The viz_ scripts will help you visualize the results.

So, for example execute:

python viz_progress.py <path/to/experiment/results>

or

python viz_progress.py <path/to/experiment/results>