Simple Parallel RL Framework

A simple parallel RL framework implemented in PyTorch and OpenAI Gym for an Inverted Pendulum with image data

The default number of parallel threads is set to 1 because the percentage of the sampling period in all traning time is very small.

In other words, the sampling costs very little time, and hence the optimization space is very narrow even by employing parallel coding technique.

Always up for a chat -- shoot me an email (wgj@buaa.edu.cn) if you'd like to discuss anything.

Installing Dependency:

You can use anaconda to create a python3 environment:

conda env create -f environment.yml

If some error messages from Anaconda are raised, you could choose to install the required python3 package manually. Run the following command with CMD in Windows or Shell in Linux or MacOS:

pip3 install pytorch pygame gym opencv_python 

How to use

Enter the DQN directory, and run the python3 command 'python3 train.py --algo ddpg(dqn)':

cd Parallel-RL-framework # 
python3 train.py --algo ddpg(dqn)

When testing the bulit environment, you could let the code idle with the following command:

python3 train.py --idling

When you run these codes, it can automatically create two subdirectories under the current directory: ./data/model/ & ./data/plots/. These two directories are used to store the models and the results.

After training, you can plot the results by running result_show.py with appropriate parameters.

References:

1. Human-level control through deep reinforcement learning (DQN) [Paper]
2. Continuous control with deep reinforcement learning (DDPG) [Paper]