Implementation of a model-free reinforcement learning approach to dry stacking with irregular rocks. Project developed to support the MSc thesis From rocks to walls: a machine learning approach for lunar base construction. Follow the link for further informations.
The contents of this project are divided in three main sub-packages:
stackrl.envscontains the implementation of the simulated environment, along with related utilities;stackrl.netscontains the implementation of the neural networks used as value estimators;stackrl.agentscontains the implementation of the reinforcement learnig algorithm (DQN) used to learn.
The class stackrl.Training provides an interface for the training sessions using elements from the above sub-packages, and saves checkpoints and logs for the learning curves. Under stackrl.train you can find other utilities to load learned policies and plot the learing curves.
You can install stackrl with:
git clone https://github.com/menezesandre/stackrl.git
pip install -e ./stackrl
Alternatively, you can directly use the Docker image.
Run this package with:
python -m stackrl <command>
You can use this package via the Docker image:
docker run --rm -u $(id -u):$(id -g) -v $(pwd):\home -v \home menezesandre/stackrl <command>