RLPy

RLPy

The Reinforcement Learning Library for Education and Research

Table Of Contents

Previous topic

Overview

Next topic

Getting Started

This Page

Installation

The installation consists of three steps:

  1. Download the package
  2. Install required libraries
  3. Compile C-Extenstions of RLPy

1. Download

Stable Version

You can download the latest stable version (1.2.1) of RLPy from http://acl.mit.edu/rlpy/rlpy-1.2.1.zip. Extract the package in your desired location.

Development Version

The development is maintained bitbucket at https://bitbucket.org/rlpy/rlpy. The git-repository with the latest development version can be cloned via:

git clone https://bitbucket.org/rlpy/rlpy.git RLPy

This will give you a copy of the repository in the directory RLPy. You might want to change the location as you wish.

2. Dependencies

RLPy requires the following packages besides Python:

GCC >= 4.6
for compiling some C++ extensions which use the C++11 standard.
Graphviz (optional)
for creating the graphical ouput of the code profiling tool.
Tk
as a backend for matplotlib and for visualizations of some domains.

In addition, RLPy requires Python 2.7 to run. We do not support Python 3 at the moment since most scientific libraries still require Python 2. The following Python packages need to be available:

  • Numpy >= 1.7
  • Scipy
  • Matplotlib >= 1.2
  • Networkx
  • PyTk
  • scikit-learn
  • Cython
  • joblib
  • hyperopt

Ubuntu / Debian

Install the non-python dependencies with:

sudo apt-get install graphviz tk blt tcl gcc g++

To install the Python packages we recommend using Anaconda as described in the following section . This will ensure you have the latest versions of each package.

Alternatively you can install the python packages via apt. Note however that these packages will usually be older. You can install them by executing:

sudo apt-get install python-dev python-setuptools python-sklearn python-numpy python-scipy python-matplotlib python-networkx graphviz python-pip tcl-dev tk-dev python-tk cython
pip install joblib hyperopt pymongo

OS X

Please install XCode and command line tools from the App Store or the Apple Developer Account direct download if you do not have already. These packages contain the compilers necessary for building RLPy’s C++ extensions.

You can verify that your compiler is new enough by executing:

clang++ --version

You should see somewhere in the output (based on LLVM 3.2svn) or a higher version number.

To install the Python packages we recommend Anaconda. This will ensure you have the latest versions of each package. Follow the instructions in the Anaconda section.

Windows

We recommend using Anaconda for installing Python and all dependencies. Follow the instructions in the following section. On Windows, Anaconda also comes with a gcc compiler.

Warning

Unfortunately, matplotlib shipped with Anaconda does not contain the tkagg backend, which we use by default. At the moment you need to install matplotlib manually with tkinter support for RLPy to work properly. We hope this issue is fixed soon. See also https://groups.google.com/a/continuum.io/forum/#!topic/anaconda/G4McL1bclAs for updates.

If you see an error complaining that the module _tkagg could not be imported, change the matplotlib_backend variable in Tools.GeneralTools to “qtagg”. While this workaround allows you to use matplotlib, it may result in interactive matplotlib plots to not be shown.

Warning

A couple of problems arise when building our Cython / C++ Extensions on Windows. It requires therefore some workarounds to get all extensions running on Windows. For details see https://bitbucket.org/rlpy/rlpy/issue/31/windows-anaconda-installation-problems Unfortunately, the problems are caused by packages we rely on and are therefore not easy to resolve for us.

Anaconda

We recommend using the Anaconda Python distribution. This software package comes with a current version of Python and many libraries necessary for scientific computing. It simplifies installing and updating Python libraries significantly on Windows, MacOS and Linux. Please follow the original installation instructions of Anaconda.

After installing Anaconda, install the dependencies of RLPy by executing:

conda install numpy scipy matplotlib networkx tk scikit-learn cython
pip install joblib hyperopt pymongo

3. Build Extensions of RLPy

Build the C++ / Cython extensions of RLPy by executing in your RLPy directory:

python setup.py build_ext --inplace

Note

If you are using a MacOS and a MacPorts version of gcc and you get an error about the -arch parameters, try using:

ARCHFLAGS="" python setup.py build_ext --inplace

Note

If you use Anaconda and get an error about incompatibility to the deployment target try instead:

MACOSX_DEPLOYMENT_TARGET=10.7 python setup.py build_ext --inplace

Tip

You can verify that your rlpy installation works well by running the testsuite in the tests directory. You can do so by executing from the rlpy directory:

nosetests tests --exe

RLPy is now successfully installed. For an introduction on how to use the framework have a look at Getting Started.