rlworkbench

RL Workbench - A comprehensive, interactive framework for RL experimentation

The code consists of two main parts:

• The python-backend is a Python webserver and collection of tools which provides functionality to setup and explore training environments, initiate training runs and benchmark trained agents. It serves functionality at a routing API or via standalone modules such as the trainer, which can be called from the command line to initiate training runs.
• The react-frontend directory contains the code for a Javascript/React-based webtool, which serves as a frontend provided to the experiment_server backend.

You can run a docker-compose stack by running docker-compose up. You do not need to install the requirements in this case.

Experiment Server

Make sure to have have git lfs intalled (e.g. via brew install git-lfs, git install lfsand then use git lfs pull):

All requirements can be installed via

pip install -r requirements.txt

We recommend the use of virtual env or similar. To use the gpu-enabled version of PyTorch, you have to install PyTorch via the official website https://pytorch.org/get-started/locally/.

Web Application

The web application is served via the experiment server. A Javascript distributable bundle is already available in the repository, and should run out of the box. If you want to build the web application yourself:

1. Make sure to have npm,yarn or an equivalent package manager installed. Switch to the tool directory via cd tool and run npm install to install all packages specified in package.json.
2. Now run either npm run build to create a production build or npm run dev-build for a debug build.
3. Restart the experiment server if the the build ran sucessfully to serve the new bundle.

Run the Application

Run the application by using docker and the given docker-compose.yml. Alternative you can start the experiment server manually by running python app.py and the frontend by running npm start.

You can access the tool via the following URL: http://localhost:8080/.

Disclaimer

The experiment server was tested on both Windows 10 and macOS. The web application was tested both on Chrome and Firefox.

This project is still in its prototype stage and may contain bugs. Be aware, that specific steps, e.g. the benchmarking of trained agents for many steps may take a lot of time. Raise your questions to @ymetz.