This repository is meant as a very minimal implementation of DRL methods to be used as a starting point for further experimentation. If you have fixes, or changes to the default parameters that will make the methods work better, please create a pull request or tell me. To use this code as a starting point, either fork the repo or download a local copy.
This code makes use of the CoR control benchmarks, a set of control problems where the state of a dynamical system should be driven towards a reference (goal) state. Some of these benchmarks have physical counterparts within the Cognitive Robotics department of the Delft University of Technology.
The code requires python 3.6+ with either tensorflow or tensorflow-gpu installed:
For the CPU version of tensorflow:
pip install tensorflow
For the GPU version:
pip install tensorflow-gpu
The remaining requirements can be installed by running the following command from the main repository directory (after downloading / cloning):
pip install -r requirements.txt
The main directory contains the simple_run.py file which currently includes a basic (and not overly effective) implementation of the Normalized Advantage Functions algorithm on the Magman benchmark. It does include comments on things that can be improved, so good luck!