Keynote: AlphaStar: How Google Infrastructure Made it Possible
summary
AlphaStar is a full StarCraft II agent built on Google-scale infrastructure that reached Grandmaster in human ladder play with no game restrictions. The project uses massive distributed simulation and training — running roughly one million games across a population-based training of agents over weeks — plus ablation experiments and latency/observation engineering to shrink action space and meet millisecond constraints. The system is productionized, continuously self-played, and demonstrably beats top human players.
AI-written summary · about
chapters (8)
- 0:00Introduction to AlphaStar
- 2:17Why StarCraft is complex
- 4:28Evolution of AI in games
- 7:39AlphaStar’s technical approach
- 10:56Training and competition details
- 13:49League system and player dynamics
- 16:22Infrastructure and scalability
- 18:32Q&A and closing
related talks
55mTearing down AlphaStar
56mCan I do AI?
01h 03mIEEE CoG 2019 - Industry Day - David Silver (DeepMind)
24mApplied AI research with StarCraft, Torch and the BEAM
40mWhat did AlphaGo do to beat the strongest human Go player?
37mA Funny Thing Happened On The Way To Reimplementing AlphaGo
14mAI for Science
28mScaling AI Infrastructure