Evolved Policy Gradients
OpenAI's metalearning breakthrough: agents that generalize to tasks they've never seen before.
Why it matters
Evolved Policy Gradients demonstrates a new approach to training efficiency and task generalization—agents can succeed on novel tasks outside their training distribution, which matters for building more adaptable AI systems.
The key facts
10 to knowOpenAI released Evolved Policy Gradients (EPG), a metalearning approach
EPG evolves the loss function of learning agents
Enables fast training on novel tasks
Agents tested on navigation tasks outside training regime (different object placement)
Demonstrates out-of-distribution generalization capability
Published April 18, 2018
Metalearning approach evolves loss functions of learning agents
Agents generalize to out-of-distribution scenarios (e.g., navigating to object on different side of room)
Published April 18, 2018 — historical research release
OpenAI research contribution
Go to the source
OpenAI Blogopenai.com
Publisher excerpt: We’re releasing an experimental metalearning approach called Evolved Policy Gradients, a method that evolves the loss function of learning agents, which can enable fast training on novel tasks. Agents trained with EPG can succeed at basic tasks at test time that were outside their training regime,…