====== Deep Deterministic Policy Gradient on PyTorch

Overview

The is the implementation of Deep Deterministic Policy Gradient <https://arxiv.org/abs/1509.02971> (DDPG) using PyTorch <https://github.com/pytorch/pytorch>. Part of the utilities functions such as replay buffer and random process are from keras-rl <https://github.com/matthiasplappert/keras-rl>_ repo. Contributes are very welcome.

Dependencies

• Python 3.4
• PyTorch 0.1.9
• OpenAI Gym <https://github.com/openai/gym>_

Run

• Training : results of two environment and their training curves:

  • Pendulum-v0

  .. code-block:: console

  $ ./main.py --debug

  .. image:: output/Pendulum-v0-run0/validate_reward.png :width: 800px :align: left :height: 600px :alt: alternate text

  • MountainCarContinuous-v0

  .. code-block:: console

  $ ./main.py --env MountainCarContinuous-v0 --validate_episodes 100 --max_episode_length 2500 --ou_sigma 0.5 --debug

  .. image:: output/MountainCarContinuous-v0-run0/validate_reward.png :width: 800px :align: left :height: 600px :alt: alternate text

• Testing :

.. code-block:: console

$ ./main.py --mode test --debug

TODO