Rainbow

Why we included this project

Rainbow is a compact PyTorch codebase that bundles the algorithmic extensions from the Rainbow paper into one configurable script: double learning, prioritised replay, dueling networks, multi-step returns, distributional Q-learning, and noisy nets. It runs out of the box on Atari environments, and pretrained models ship with the releases, so you can sanity-check your own training runs against published results. Researchers or engineers reproducing numbers from the paper will find the flags for data-efficient training and alternate architectures handy for comparing variants. It is a research tool rather than a production service, but it works well as a compact template for building reinforcement learning pipelines around game environments.