Video games such as Dota 2 serve as a natural testbed for research in training artificial intelligence agents, providing ready-made simulated worlds with significant depth and complexity as well as large communities of human experts to learn from and benchmark against. Last year, OpenAI Five became the first artificial agent to defeat the reigning world champion of a top professional e-sport.
In this talk I’ll tell the story of the Dota project at OpenAI, explain some lessons we learned, and situate these results relative to the recent dramatic progress of intelligent agents interacting in complex systems.