This is an implementation of Deep Deterministic Policy Gradient (DDPG) in bullet Gym using PyTorch.

• Python 3.6
• PyTorch 0.3.0
• openAI Gym
• pybullet

• here is a simple example to train CartPole with high efficiency:

$ python main.py --debug --discrete --env=CartPole-v0 --vis --tau=0.02 --discount=0.97 --warmup=1000 --window_length=2

• the results after 1 min training:

#240: episode_reward:129.00 steps:1107 noise:0.28 time:4.10,3.15
#241: episode_reward:132.00 steps:1140 noise:0.24 time:2.14,3.17
#242: episode_reward:107.00 steps:1167 noise:0.12 time:1.73,3.24
#243: episode_reward:63.00 steps:1183 noise:0.21 time:1.00,3.21
#244: episode_reward:117.00 steps:1213 noise:0.24 time:1.90,3.16
#245: episode_reward:140.00 steps:1248 noise:0.14 time:2.27,3.82
#246: episode_reward:73.00 steps:1267 noise:0.45 time:0.09,3.41
#247: episode_reward:194.00 steps:1316 noise:0.08 time:3.19,3.21
#248: episode_reward:306.00 steps:1393 noise:0.14 time:5.05,3.18
#249: episode_reward:199.00 steps:1443 noise:0.03 time:3.27,3.14
[Evaluate] Step_0001463: mean_reward:343.0 and save model

• you can using this to understand each arguments' usage

$ python main.py --help