PyPlan is an easily-extensible library of planning algorithms and environments, with OpenAI Gym integration for environments with discrete action spaces.
Running a one-step greedy bandit algorithm on Berkeley's Pacman simulator is as simple as:
u_ro = uniform_rollout_agent.UniformRolloutAgent(depth=0, num_pulls=100)
pacman = pacman_dealer.Dealer(layout_representation='originalClassic')
pacman.run(agents=[u_ro], num_trials=10)
To implement your own agents or environments, consult the documentation in abstract/. A tutorial series based on this suite is currently in production.
After cloning the project, get started with the files in the demos/ directory.
Unix and Mac users must follow the OpenAI installation instructions. If you get runtime errors, you may need to replace the installed gym folder with WinPython/python-3.6.1.amd64/Lib/site-packages/gym.
Although OpenAI does not officially support Windows, Windows users may use the included WinPython interpreter (located in the WinPython/python-3.6.1.amd64/ subdirectory) to access the Gym's full functionality. Furthermore, to record episodes for games with video output, Windows users must install ffmpeg - make sure to add the bin subdirectory to your system path!