GitLab

Improve the consitency across classes when verbose mode is on by using a
single function that does proper formatting.

fixed firemdp.py AssertionError caused by incorrect argument sequence

Make the code more compliant to PEP 8.

Link to the modules, classes or functions that each module provides in
the module level docstring.

Add skip_check argument in MDP and subclasses

Allowing to skip checking the transition and reward matrices on creation
of an MDP instance (and all of its subclasses).

Separate downloading from PyPI and GitHub into separate sections. Be
more explicit that the package is on PyPI, and use the PyPI shields.

Attempt at creating a better tagging and release process.

The linear programming algorithm provides incorrect results. Make this a
'private' class of the ``mdp`` module until until it can b e made to
work correctly.

Sparse rewards were broken. This commit adds several more functions to
MDP class to make the reward computation code more modular and hopefully
more correct. Rewards given as sparse matrices are converted to a dense
vector. Future work will ensure that rewwards gieven in sparse format
remain sparse. Fixes #7.

Finite horizon tests are commented out until the horizon can be passed.

Use Python code block directive for the example usage. Link to the license file.

Seperate out functionality into more modular and single purpose
functions. Move exception classes to own module for this purpose. Add
new user facing functions for checking MDPs.

Add two new functions _randDense and _randSparse to separate
functionality from the rand function. Call these functions from rand.

Base the comparison that two arrays are almost equal based on the
spacing of the maximum value of the array rather than one. Using one
will give incorrect resulsts when the magintude of the values is
significantly different to one.

Two unit tests based on the small example for the discounted and
undiscounted case.

Add a base class that other classes for testing the MDP algorithms can
use it.

Add an example usage for the small example to the docstring for test
coverage purposes.

Add a small MDP example that has only two states and two actions. This
is mostly useful to help with testing the algorithms, and as an example
of the usage of them.

Fix V initialization in PolicyIterationModified

V was initialized with the wrong dimensions for undiscounted MDPs,
which caused an exception to be raised in _bellmanOperator.

Use ellipsis so that Python 2 and Python 3 outputs will pass. Check that
floating point numbers are within 1e-12 of the expected answer rather
than exact matches.

The CVXOPT python package has some optional dependencies that must be
altered in the setup.py file for support to be compiled in. GLPK is one
if these. At the moment the setup.py script is not being altered while
installing dependencies, so downloading cvxopts extra dependencies is
useless.

Move the test command from .travis.yml to runtests.sh to allow easy
testing on workstations.