This is a Tensorflow-based implementation of Trust Region Policy Optimization, a deep reinforcement learning technique. This code has been tested on Python 3.5+.
This implementation is intended to be more pedagogical than optimal. We focus here on understanding the algorithm via clear code. That said, when run on a GPU, it performs just fine on moderately large problems.
This code is associated with a forthcoming blog post detailing the algorithm and discussing how to efficiently implement it using tensorflow. Check back here for that link.
Installation is easy. Just add the root directory to your local Python path in
your .profile.
export PYTHONPATH="${PYTHONPATH}:/Users/username/path/to/dopamine"
To get started, run one of the simple experiments in the /scripts directory.
I recommend starting with /scripts/lineworld_experiments.py to see get a
quick sense of how this works.
Tests can be executed using nose. Just
run it from the tests/ directory.