reaver

Why we included this project

Reaver keeps the agent, model, and environment as separate components you can swap out, which makes it a practical framework for deep reinforcement learning experiments, particularly around StarCraft II. It also supports Gym, Atari, and MuJoCo environments. The included command-line tools and pre-tuned hyperparameters let you start a training run with a short one-liner and watch it move toward the reference results, so it works as a hands-on reference for learning how actor-critic methods behave in practice. Hobbyists can get a working agent by tweaking a small set of hyperparameters, while researchers get a codebase they can extend. One caveat: the project is no longer maintained, so treat it as a study resource and a starting point rather than a dependency for a long-lived production system.