MiniDRL

Minimal implementations of distributed deep reinforcement learning algorithms, with a focus on recurrent neural networks. Heavily inspired by CleanRL and CORL this library provides high-quality and easy-to-follow stand-alone implementations of some distributed RL algorithms.

Getting Started

Prerequisites:

• python >= 3.10 (tested with 3.10)

To install:

git clone [email protected]:Jjschwartz/miniDRL.git
cd miniDRL
pip install -e .
# or to install all dependencies
pip install -e .[all]

Run PPO on gymnasium CartPole-v1 environment using four parallel workers (reduce number of workers if you have less than four cores, or feel free to increase it if you have more):

python minidrl/ppo/run_gym.py \
    --env_id CartPole-v1 \
    --total_timesteps 1000000 \
    --num_workers 4

# open another terminal and run tensorboard from repo root directory
tensorboard --logdir runs

To use experiment tracking with wandb, run:

wandb login  # only required for the first time
python minidrl/ppo/run_gym.py \
    --env_id CartPole-v1 \
    --total_timesteps 1000000 \
    --num_workers 4 \
    --track_wandb \
    --wandb_project minidrltest

Algorithms

This repository contains standalone implementations of some of the main distributed RL algorithms that support recurrent neural networks, including:

PPO - Single Machine

paper | code | docs

Learning curve of PPO - Single Machine on Atari Pong with different number of parallel workers

R2D2

paper | code | docs

Maybe in the future

• PPO - Multi Machine
• IMPALA
• R2D2 - Multi Machine