This is an demonstration of the PI^2 (Policy Improvement with Path Integrals) algorithm implemented in Python with computations done on a GPU through PyCuda. The program learns a set of controls that guides a virtual flying object through a tunnel without crashing into the floor or ceiling. A few examples can be run on any computer with python and pygame. These can be found in the flybot file, a computer with PyCuda enabled can run any arbitrary tunnel by creating a tunnel and feeding it into the flybot.py program. PI2.py is the python/cuda program which handles all of the PI^2 algorithm.
For more information about this project, or to get in contact with it's creator please contact:
Grady Williams gradyrw@gmail.com