Thread by @DeepMind

Reinforcement learning typically trains & tests agents on the same game. New work shows how our team trains generally capable agents on huge game spaces, resulting in agents that generalise to held-out test games, & learn behaviours like experimentation dpmd.ai/open-ended-blog 1/

Rather than training on a limited number of tasks, our team defines a whole universe of tasks that can be procedurally generated, from simple object finding games to complex strategic games like Capture the Flag. 2/

By constructing a hierarchical learning process with an open-ended and iteratively refined objective, it was possible to train agents that never stop learning, and develop increasingly general behaviour across games. 3/

Our team finds these agents are able to generalise to many hand-authored probe tasks, and can solve out-of-distribution probe tasks through experimentation and tool use. 4/

Paper and results video below:
dpmd.ai/open-ended-pap…
dpmd.ai/open-ended-vid…

By the Open-Ended Learning Team including @maxjaderberg and @wojczarnecki.

Interested in this work? A research scientist role on the team is currently open, learn more: dpmd.ai/OE_RS 5/5