/DDPG-CARTPOLE

Stable and robust control a cartpole with DDPG in continuous actions

Primary LanguagePython

DDPG_CARTPOLE

  • Stable and robust control a cartpole in continuous actions with large noise by using DDPG.

Environment Description

  • We use OpenAI's cartpole, but make its actions continuous.
  • And there are many noise in this environment setting, but our policy is still very robust.

Internal uncertainty

  • In every 0.02s, the Cart's mass changes in a gaussian distribution (1,0.2).
  • In every 0.02s, the Pole's mass changes in a gaussian distribution (0.1,0.02).
  • In every 0.02s, the gravity changes in a gaussian distribution (10,2).

Action uncertainty

  • And the action the agent chooses will also be added with a gaussian distribution(action,10).
  • The torch, the acceleration，angular acceleration all add with a gaussian distribution.

Model

  • cartploe_normal.ckpt train with no uncertainty.
  • cartploe_plus_5.ckpt train with full uncertainty.

Env

  • cartpole_env.py is without uncertainty environment.
  • cartpole_plus.py is the uncertainty environment.

Dependencies

  • Tensorflow (1.9.0)
  • OpenAi gym (0.10.8)

Reference

[1] Reinforcement-learning-with-tensorflow