policyopt_torch

This contains an implementation of ARM for discrete action space envs (https://arxiv.org/abs/1710.11424v2).

Requirements

Tested with python 2.7 and 3.5 (on ubuntu 16.04 + cuda 8.0 via nvidia-docker).

• gym commit 0c91364cd4a7ea70f242a28b85c3aea2d74aa35a
• numpy 1.13, 1.14 or newer
• opencv-python
• pytorch 0.3.1

Atari

Run python ./train_atari_arm.py with the gym Atari envs installed. See the comments in train_atari_arm.py for the various options.

Doom

Similarly, run python ./train_doom_arm.py. ViZDoom experiments use slightly customized versions of doom-py and the envs by @ppaquette:

https://github.com/peterhj/doom-py/tree/peterhj-depth
https://github.com/peterhj/gym-doom/tree/peterhj-rllab