Library contains various classes served as interface between Gym and PyBrain libraries. It allows to use PyBrain reinforcement learning algorithms on Gym's environments. In addition it allows to execute experiments in parallel fashion reducing learning time.

• execution of Gym environments in PyBrain,
• execution of multiple experiments in parallel manner,
• custom transformation of data passed between Gym and PyBrain,
• quantization/digitization of continuous space (floating point) values and arrays of values to discrete number of states (integers),

In directory src/examples there are following examples:

• frozenlake/frozen.py -- solution for FrozenLake-v0 problem using SARSA learner.
• cartpole/cart.py -- solution for CartPole-v1 problem using SARSA learner. It demonstrates how to discretize continuous input.
• mountaincar -- solution for MountainCar-v0 and MountainCarContinuous-v0 problem using SARSA learner. In addition, continuous version run experiments in parallel manner.
• multiprocess -- example of use of multiprocessing library.

1. checkout the project
2. execute installenv.sh script to install/configure virtual environment
3. execute startenv.sh to activate virtual environemnt
4. execute installpackage.sh to install pybraingym library in development mode
5. execute src/runtests.py to run unit tests

Library requires fixed version of PyBrain library to work. It can be checked out from following address: https://github.com/anetczuk/pybrain

Unit testsing runner script supports code profiling and code coverage options. Run script with --help argument for more info.

Directory tools contains following scripts:

• notrailingwhitespaces.sh -- removes trailing whitespaces from py files
• converteol.sh -- removes ^M character (Windows) from py files
• codecheck.sh -- run pep8 and flake8 checks against source code
• doccheck.sh -- run docstyle against source code

Q/SARSA learner does not handle well last reward of episodic environments causing last reward and state to be ignored. It can be crucial in case when non-zero reward is only received when reaching goal. As workaround, processLastReward function was introduced to call experiment fake iteration to allow agent to store the data.