Deep-Reinforcement-Learning-Hands-On-Second-Edition, published by Packt

Modified by Simon D. Levy:

1. Made a proper Python package (currently Chapter 19 only)

2. Removed RoboSchool dependency.

3. Support one-dimensional action space.

% cd drlho2e/ch19
% python3 trpo.py -e Pendulum-v0 -n pendulum

After a hundred thousand iterations or so, the program should should report saving the current best agent to a file. After a few million iterations, the best agent should be good enough to test, which you can do as follows:

% python3 play.py --render -e Pendulum-v0 -m saves/trpo-pendulum/best_-<REWARD>_<ITER>.dat

where <REWARD> is the amount of reward and <ITER> is the number of iterations at which it was saved. (It is easiest to do this through tab completion.) You should see brief animation of the pendulum rising to an upright position, indicating success.

To work with your own Gym environment (like this one), you'll want to install this repository as a Python package. Back in the top-level directory:

% python3 setup.py install

On Linux and OS X, you'll likely have to do this with sudo:

% sudo python3 setup.py install

Here are some examples of simple programs I wrote using this package with my own Gym envrionment.