DDPG

Implimentation of DDPG algorithm which is installable with pip.

The original DDPG algorithm was proposed in the paper: Continuous Control with Deep Reinforcement Learning

http://arxiv.org/abs/1509.02971

It is still a problem to implement Batch Normalization on the critic network. However the actor network works well with Batch Normalization.

Installation

Install the package

pip3 install ddpg

Getting Started

The DDPG algorithm acts on environments which follow the openai-gym api.

# Create a test environment with gym
env = gym.make('MountainCarContinuous-v0')

Train the DDPG agent:

from ddpg import DDPG

# Create a new agent
agent = DDPG(env)

# Train the agent
agent.train()

# Save the weights
agent.model_save()

Dependencies

• Python3
• Tensorflow 1.1
• NumPy
• Matplotlib

Some Evaluations

1 InvertedPendulum

2 InvertedDoublePendulum

3 Hopper unsolved

Reference

1 https://github.com/rllab/rllab

2 https://github.com/MOCR/DDPG

3 https://github.com/SimonRamstedt/ddpg