DDPG

Reimplementing DDPG from Continuous Control with Deep Reinforcement Learning based on OpenAI Gym and Tensorflow

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.

Some Mujoco environments are still unsolved on OpenAI Gym.

Some Evaluations

1 InvertedPendulum

2 InvertedDoublePendulum

3 Hopper unsolved

How to use

git clone https://github.com/songrotek/DDPG.git
cd DDPG
python gym_ddpg.py

If you want to change the Gym environment, change ENV_NAME in gym_ddpg.py.

If you want to change the Network type, change import in ddpg.py such as

from actor_network_bn import ActorNetwork
to
from actor_network import ActorNetwork

Reference

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

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

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