home
past talks
resources
conferences
|
|
|
Acquiring useful skills in reinforcement learning
Özgür Şimşek
UMass
Abstract
What is a useful skill for an autonomous agent? We provide one possible answer to this question using the concept of "access states" -- states that allow the agent to transition to a different part of the state space. We present two different methods for identifying access states -- one based on a measure of relative novelty, and another based on local graph partitioning -- and illustrate the utility of these methods in a number of reinforcement learning problems.
Back to ML Lunch home
|