Reproduce “SCC: an efficient deep reinforcement learning agent mastering the game of StarCraft II” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiangjun Wang, Junxiao Song, Penghui Qi, Peng Peng, Zhenkun Tang, Wei Zhang, Weimin Li, Xiongjun Pi and 4 more

StarCraft Commander (SCC), developed by inspir.ai, achieved top GrandMaster level performance in the full game of StarCraft II. It accomplished this feat with orders of magnitude less computational resources than previous state-of-the-art DRL agents, notably by using around 16 times fewer concurrent game environments during training.

Key insights

  • Achieving top human-level performance in StarCraft II is possible with significantly reduced computational resources, demonstrating that architectural and algorithmic efficiency can mitigate the need for extreme scale.
  • High-quality, focused datasets for imitation learning, even when smaller, are crucial for effective initial policy acquisition, especially when followed by targeted fine-tuning.
  • A combination of network pruning, specialized attention mechanisms, and strategic reward shaping in reinforcement learning leads to a more efficient and generalizable agent that can also discover novel strategies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.