Inverse Reinforcement Learning (IRL) Augmentation for Markov Decision Process (MDP) Toolbox for Python
The MDP toolbox for Python by Sam Cordwell provides classes and functions for the resolution of descrete-time Markov Decision Processes. The list of algorithms that have been implemented includes backwards induction, linear programming, policy iteration, q-learning and value iteration along with several variations.
This IRL augmentation for this toolbox will provide additional classes and functions related to infinite time horizon maximum causal entropy IRL. The list of algorithms for deriving soft Bellman policies that will be implemented includes soft value iteration, soft q-learning, and convex optimization. These algorithms are described in 'Infinite time horizon maximum causal entropy inverse reinforcement learning' by Michael Bloem and Nicholas Bambos, IEEE Conference on Decision and Control, Los Angeles, CA, 2014 doi 10.1109/CDC.2014.7040156
Features under development
- Three algorithms for deriving soft Bellman policies
- Three algorithms for policy evaluation
- Gradient algorithm for tuning parameters in soft Bellman policies
PLEASE NOTE: the linear programming algorithm is currently unavailable except for testing purposes due to incorrect behaviour.
Installation
NumPy and SciPy must be on your system to use this toolbox. Please have a look at their documentation to get them installed. If you are installing onto Ubuntu or Debian and using Python 2 then this will pull in all the dependencies:
sudo apt-get install python-numpy python-scipy python-cvxopt
On the other hand, if you are using Python 3 then cvxopt will have to be compiled (pip will do it automatically). To get NumPy, SciPy and all the dependencies to have a fully featured cvxopt then run:
sudo apt-get install python3-numpy python3-scipy liblapack-dev libatlas-base-dev libgsl0-dev fftw-dev libglpk-dev libdsdp-dev
The main way to download the package is from GitHub.
To clone the Git repository
git clone https://github.com/nasa/pymdptoolbox.git
Change to the PyMDPtoolbox directory
cd pymdptoolbox
Install via Setuptools, either to the root filesystem or to your home directory if you don't have administrative access.
python setup.py install
python setup.py install --user
Read the Setuptools documentation for more advanced information.
To learn how to use Git then you might consider reading the freely available Pro Git book written by Scott Chacon and Ben Straub and published by Apress.
Quick Use
Start Python in your favorite way. The following example shows you how to import the module, set up an example Markov decision problem using a discount value of 0.9, solve it using the value iteration algorithm, and then check the optimal policy.
import mdptoolbox.example
P, R = mdptoolbox.example.forest()
vi = mdptoolbox.mdp.ValueIteration(P, R, 0.9)
vi.run()
vi.policy # result is (0, 0, 0)
Documentation
Documentation is available as docstrings in the module code.
If you use IPython to work with the toolbox,
then you can view the docstrings by using a question mark ?
. For example:
import mdptoolbox
mdptoolbox?<ENTER>
mdptoolbox.mdp?<ENTER>
mdptoolbox.mdp.ValueIteration?<ENTER>
will display the relevant documentation.
Contribute
Source Code: https://github.com/nasa/pymdptoolbox
License
See LICENSE.txt for details.