Applied AI research with StarCraft, Torch and the BEAM
summary
A community bot system implements a blackboard message bus that extracts live StarCraft events, stores them on the BEAM and feeds Python modules for model training with Torch/PyTorch. Implementation uses Luerl on the BEAM for coordination, TorchCraft + 0MQ to glue the StarCraft engine to backends, and a distributed key‑value store for state. The system runs heuristics in LuaJIT, trains neural nets offline, and is functional but not yet battle‑tested in competition.
AI-written summary · about
chapters (10)
- 0:00Start
- 1:36Why Do We Choose Starcraft
- 3:14Jean Piaget Definitions of Intelligence
- 4:09Intellectual Adaptation
- 4:45Organizational Behavior
- 6:52Internal Consistency
- 7:50Progression of Goals
- 8:29The Blackbird Model
- 10:38Why Do We Use the Blackboard Approach
- 21:03How Did You Train the Model for Your Inference
related talks
28mSAIDA RL: An Open-source Reinforcement Learning Platform
19mKeynote: AlphaStar: How Google Infrastructure Made it Possible
20mEngaging Bots for Multiplayer Games
55mTearing down AlphaStar
18mBuilding AI Agents
01h 10mReal-time AI Lakehouse
36mMachine Learning with Apache Beam
01h 01mIntroduction to Practical AI