DI-star

Why we included this project

DI-star is a good place to start if you want to see how a game-playing agent actually gets trained to a high level. Its pipeline has two stages: supervised learning on recorded human matches, then reinforcement learning with self-play and league-style matches, and that combination produced a grandmaster-level Zerg agent. The platform is built for large-scale distributed training, but the authors also ship guidance for running a meaningful training baseline on a single PC, so you do not need a cluster just to experiment. Researchers and RL engineers get training code, pretrained SL and RL checkpoints, and a playable demo to build on or reproduce. Just know it is a StarCraft II-specific system rather than a general RL library, so expect to spend some time on SC2 tooling and environment setup before anything runs.