This distributed framework supports the Proximal Policy Optimization (PPO) algorithm along with off-policy correction methods like V-trace and TD(λ). The experiences collected by multiple actors are sent to a single buffer and later on being sampled by the learner for minibatches updates. The actors sync its weights periodically with the learner.

Inter-process communication (IPC) between actors and the learner is implemented by socket programming, which provides a simple network interface for IPC between threads, processes and even different machines over the Internet.

Clone this repository, create a Python virtual environment and run pip3 install -r requirements.txt.

In terminal

TIME=15 NWORKERS=3 make run_lunar

This means that we will train an agent for the task run_lunar with 3 workers and the training process will shutdown automatically after 15 minutes. For more details, check Makefile.

We aim to make this framework highly modular and flexible to any algorithms, environments and design choices. For example, we can add custom neural network architectures into the networks folder and quickly adapts to new sets of hyperparamters by modifying or adding config files.

Because we are still developing this framework, feel free to make suggestions and report bugs.

Future plan:

• Add UPGO.
• A wrapper for custom environments aparts from OpenAI gym.
• A wrapper for custom choices of networks (like shared network for Actor-Critic, ConvNet-based networks, ...).
• PopArt.