reaver  by inoryy

Deep RL framework for StarCraft II tasks (Gym, Atari, MuJoCo also supported)

Created 8 years ago
560 stars

Top 57.3% on SourcePulse

GitHubView on GitHub
Project Summary

Reaver is a modular deep reinforcement learning framework designed for StarCraft II tasks, offering a flexible and performant solution for researchers and hobbyists. It aims to replicate DeepMind's state-of-the-art results in complex game environments, while also supporting popular benchmarks like Atari and MuJoCo.

How It Works

Reaver employs a modular architecture, decoupling agents, models, and environments for easy swapping and extension. It leverages shared memory and lock-free multiprocessing for significant performance gains (up to 1.5x in StarCraft II) on single-machine setups, a key advantage over IPC-based multiprocessing approaches. Configuration is managed via gin-config, allowing for easy hyperparameter tuning and experiment sharing.

Quick Start & Requirements

  • Install: pip install reaver[gym,atari,mujoco]
  • Requirements: PySC2 >= 3.0.0, StarCraft II >= 4.1.2, TensorFlow >= 2.0.0, TensorFlow Probability >= 0.9. Optional extras for Gym, Atari, and MuJoCo.
  • Setup: Linux recommended for performance/stability. A Google Colab notebook is available.
  • Docs: https://github.com/inoryy/reaver-pysc2

Highlighted Details

  • Achieves competitive results on StarCraft II minigames, matching DeepMind SC2LE benchmarks.
  • Optimized for single-machine multiprocessing via shared memory, offering substantial speedups.
  • Supports StarCraft II (via PySC2), OpenAI Gym, Atari, and MuJoCo environments.
  • Bundled with pre-trained weights and Tensorboard logs for reproducibility.

Maintenance & Community

  • Project Status: No longer maintained.
  • Community: SC2AI online community (Discord).
  • Contact: Email provided for questions.

Licensing & Compatibility

  • License: Not explicitly stated in the README.
  • Compatibility: Designed for research; commercial use implications are unclear due to lack of explicit licensing.

Limitations & Caveats

The project is no longer maintained, meaning no future updates or bug fixes are expected. While it supports multiple environments, the primary focus and most extensive testing appear to be on StarCraft II. The lack of a clear license may pose compatibility issues for commercial or closed-source projects.

Health Check
Last Commit

4 years ago

Responsiveness

Inactive

Pull Requests (30d)
0
Issues (30d)
0
Star History
0 stars in the last 30 days

Explore Similar Projects

Starred by Deepak Pathak Deepak Pathak(Cofounder of Skild AI; Professor at CMU), Anastasis Germanidis Anastasis Germanidis(Cofounder of Runway), and
1 more.

deer by VinF

0%
489
Deep reinforcement learning framework
Created 9 years ago
Updated 2 months ago
Starred by Evan Hubinger Evan Hubinger(Head of Alignment Stress-Testing at Anthropic), Jiayi Pan Jiayi Pan(Author of SWE-Gym; MTS at xAI), and
1 more.

rl by pytorch

0.4%
3k
PyTorch library for reinforcement learning research
Created 3 years ago
Updated 2 days ago
Starred by Alex Yu Alex Yu(Research Scientist at OpenAI; Former Cofounder of Luma AI), John Yang John Yang(Coauthor of SWE-bench, SWE-agent), and
6 more.

cleanrl by vwxyzjn

0.5%
8k
RL algorithms implementation with research-friendly features
Created 6 years ago
Updated 2 months ago
Feedback? Help us improve.