This is the reposository for Thomas Tarler's practicuum 2 in reinforcement learning.

Directory

1. Final Report: Latex documents with final results.
2. : (Note to instructor:) Use this PDF when evaluating writeup, as the ReadMe.MD is rather sparse
3. scripts: Scripts that can be run on a Python cluster to determine final results

Final Neural Networks:

1. 

2. DDPG Architecture a.

   b.

   c.

   d.

To Run Scripts

1. Zip and upload DDPG or DQN to the multiprocessing cluster of your choice. Must have the following: a. anaconda3 with keras-rl, box2d installed b. If you want to render, ensure that ffmpeg (use brew, yum, or apt depending on server) c. Tensorflow 1.0 compatability installed
2. For DQN, run python DQN_BipedalWalker.py, no command line arguments are necessary. If you do not have rendering established, put n when prompted.
3. For DDPG, run python ddpg_batch.py EPISODES BATCH_SIZE LSTM ARG EPC. Again, select n for rendering. Command line arguments are: a. EPISODES in integer is number of episodes (I ran 10000) b. BATCH_SIZEin integer is the batch size (typically 16-32) c. LSTM is a boolean, depending if you want to call a LSTM network for the recall buffer d. ARG in integer is how often you want training results displayed via CLI e. EPC in integer is how many epochs you want to use.