Multi-Agent AI Mastery of StarCraft: Key Takeaways from Alibaba’s 2017 Seminar

In April 2017, Alibaba hosted an "AI Agents" academic seminar in Beijing, gathering scholars from computer science and neuroscience to explore cross‑disciplinary advances in artificial intelligence.

The Alibaba Cognitive Computing Lab, in collaboration with University College London, used StarCraft: Brood War micro‑battles as a testbed and introduced the Bi‑directional Collaborative Network (BiCNet), which enables multiple agents to automatically learn optimal strategies—from collision‑free movement to complex cover attacks and concentrated fire.

BiCNet achieved the highest win‑rate among contemporary methods, drawing attention from top research institutions such as Oxford, KAIST, Tsinghua, and Shanghai Jiao Tong University.

Senior Director Yuan Quan explained that recent AI progress is heavily inspired by neuroscience, especially deep reinforcement learning, and highlighted challenges like reasoning under uncertainty, multi‑agent cooperation, and balancing short‑ and long‑term rewards, which are far more complex in StarCraft than in deterministic games like Go.

Yuan also linked the research to Alibaba's e‑commerce ecosystem, noting that current recommendation bots operate independently and that enabling collaborative bots could create greater value for users and sellers, with StarCraft providing an ideal simulation environment.

Professor Wang Jun (UCL), a co‑designer of BiCNet, discussed the classic deep reinforcement learning behind AlphaGo and introduced the concept of artificial collective intelligence. He also covered recent advances in GANs, including SeqGAN for sequential generation, and reported a new GAN paper achieving a score of 8.34 at ICLR.

Neuroscience speakers, including Song Sen and Wu Si, emphasized brain‑inspired computing, pointing out the brain's energy efficiency (≈10 W) compared to the massive power consumption of deep learning systems. They described advances in understanding local circuit structures, dynamic information processing, and prediction as essential for true intelligence.

The panel concluded with four key questions: why AI should be brain‑like, how to draw principles from neuroscience, the current state of neuromorphic chips, and how much of the brain we truly understand.