This repository incorporates the World Model generative model architecture developed by Ha & Schmidhuber here to enhance Deep Reinforcement Learning agents by reducing the dimensionality of the raw image states to a compressed feature vector. We consider Deep Deterministic Policy Gradients (DDPG) and Proximal Policy Optimization (PPO).
Follow these steps to install the repository dependencies
-
Clone the repository with
git clone https://github.com/shawnmanuel000/WorldModelsForDeepRL.git -
Cd into the repository root directory and install the required packages in Python3 with
pip3 install -r requirements.txt -
In case the Box2D fails to build, you may be missing 'swig'. If so, follow the instructions in ./dependencies/swig_{OS}/swig3.0.8/Doc/Manual/preface.html
-
Test the three saved models by running
python3 visualize.py
-
To train the RL agents asynchronously, run the following line with either ddpg or ppo
bash train_a3c.sh [ddpg|ppo] -
To train the RL agents synchronously, run the following line with either ddpg or ppo
python3 train_a3c.py --model [ddpg|ppo] --runs 500 -
To train the complete World Model with controller, run the following line
bash train_worldmodel.sh