Skip to Content
MIT Technology Review

Tencent’s AI programs defeat Starcraft’s own AI

AI researchers at the Chinese tech giant Tencent have posted details of two programs capable of beating the “cheating” AI found in the popular video game Starcraft.

The AI programs, called TStarBot1 and TStarBot2, were able to defeat the AI built into Starcraft, which has a superhuman view of the entire game. The programs use quite different approaches to achieve the same ends. The first employs advanced machine learning, while the second follows carefully crafted rules. Neither is sophisticated enough to defeat an expert human player, however.

Ready player none: Starcraft is an important new challenge for AI researchers, for two reasons. First, it’s a complex and sprawling game. Second, it involves a deferred reward—it’s difficult for a program to know if it’s doing well until relatively late in the game, so simple trial-and-error learning doesn’t work well.

London v. Shenzhen: The Alphabet subsidiary DeepMind, famous for developing AlphaGo, a program that mastered Go using reinforcement learning, is working on its own Starcraft AI programs. But it has yet to announce something quite as capable as Tencent’s programs.

Copycat no more: Rumor has it that Tencent has a lab dedicated to replicating DeepMind’s achievements. Yet Tencent, which operates China’s dominant mobile social app, WeChat, has been investing heavily in AI research. And it would seem that the company’s researchers are now making strides of their own.