In this talk Philip will introduce audience to high-level techniques that we have used to eventually beat best professional human players in the limited version of game Dota 2 (and continue to use to expand to full game). He will cover what is reinforcement learning (RL) and sketch out how can one apply it at scale. He will then look into the problem of exploration, in what ways current RL techniques explore solution space differently than humans would and how does self-play help to create natural curriculums. He will finish with a discussion of what are characteristics of problems where these techniques can achieve previously unseen and/or superhuman performance.
[ Ссылка ]
