ARCADES

Author: Alexis BRENON, Date: 2017-12-05

What is it?

ARCADES stands for Adaptive Reinforced Context-Aware Deep Decision System. It is, at the same time, a proof a concept of using Deep Q-Network for decision making in smart-home, and a library implementing deep reinforcement learning.

For more informations, let's refer to these publications:

• Brenon (2017). Modèle profond pour la prise de décision adaptative en habitat intelligent
• Mnih et al. (2015). Human-level control through deep reinforcement learning

How to use it?

This repository is composed of different branches. The master branch contains most of the source code:

• The arcades folder contains the code of the library.
• The assets folder contains some binary files used by the code.
• The data folder contains some real data.
• The public folder contains the generated documentation (if any).
• More folders will come with different tools like tsne or deepdream.
• Many scripts, detailed below.

The xp/* branches are related to an experiment. Most of the time they are closed with a tag (and if not, they should) allowing reproducibility of the experiments.

At the root of the repository, you can find some scripts (shell or Lua ones). Here is a brief description of each:

• install_dependencies.sh try to install all dependencies (listed in Dependencies) required
• main.lua load ARCADES and launch an experiment given its arguments. ARCADES should be updated to rely on a configuration file (JSON or so) to avoid these tricky parsing and instantiations.
• run.lua is a wrapper, which builds a well formed arguments list and finally calls main.lua.
• expe_template.lua is a template file used to create experiments. It defines parameters and then call run.lua.
• debug.lua is an instance of the expe_template, used to... Debug?
• test.lua is an instance of the expe_template used to test code (short experiment).
• cross-val.sh run multiple experiment and changing training and testing sets each time.
• cross-val-results.sh looks for experiment results in a directory and build a resulting metrics file.

Finally, the config.ld file is the configuration file of the LDoc utility used to build the documentation.

Dependencies

This project rely on dependencies installed by the install_dependencies.sh script. Some are Lua libraries, such as:

• torch: The neural network framework.
• hash: hashing functions (distributed by Torch).

Others are C libraries that you have to install yourself:

• libglib2.0-0: GLib objects management.
• libcairo2: graphics rendering (using GLib objects).
• librsvg2-2: SVG rendering.