How google-deepmind/alphastar works
Project type and purpose
• This is an open‐source deep reinforcement learning framework from DeepMind for training StarCraft II agents.
• It ships with general purpose neural network architectures (alphastar/architectures/), training/evaluation modules (alphastar/modules/), utilities (alphastar/commons/, alphastar/collections/), logging (alphastar/loggers/), and a fully offline “Unplugged” RL pipeline (alphastar/unplugged/).
File structure
• Top-level build or packaging: WORKSPACE, setup.py, Bazel (bazel/), requirements.txt.
• Core code under alphastar/ partitioned into:
• architectures/: model definitions (standard, modular, dummy) and reusable components (encoders, heads, static data, utilities).
• collections/ & types/: data structures and type definitions.
• commons/: shared utilities (logging utilities, metrics, sampling, JAX helpers).
• modules/: training and evaluation logic (agent, evaluator, match generator, optimizers, run loop).
• loggers/: episode and evaluation loggers.
• unplugged/: offline‐only pipeline (data ingestion, dataset generation, losses, learner module, training/evaluation scripts).
Inspect README for insights
• Emphasizes two main offerings: the flexible architectures and the offline “Unplugged” pipeline with Behavior Cloning.
• Describes use of PySC2 for StarCraft II environment, built via pip wheels or Bazel.
• Shows intended workflow: prepare data → run training → run evaluation.