A Pytorch implementation of Deep Determinisitc Policy Gradient for simple continuous control tasks.

More info: Continuous control with deep reinforcement learning

Watching a pretrained agent on Pendulum-v0:

python run.py --env Pendulum-v0 --agent saves/pretrained_pendulum --episodes 10

or on MountainCarContinuous-v0:

python run.py --env MountainCarContinuous-v0 --agent saves/pretrained_mountaincar --episodes 10

python ddpg.py

There are a ton of CL flags. See the bottom of ddpg.py for a full list, but here are the important ones:

• --env is the gym environment id. Options are MountainCarContinuous-v0 and Pendulum-v0
• --num_episodes is how many episodes of experience to collect during training. Defaults to 500.
• --batch_size is how many sample transitions are passed through the networks at once during training. Defaults to 128. This may need to be reduced when running on CPUs.
• --render is either 1 or 0. 1 lets you watch the agent as it learns. This slows the process down.