Structure in the space of value functions

David Foster, Peter Dayan

Research output: Contribution to journalArticlepeer-review

Abstract

Solving in an efficient manner many different optimal control tasks within the same underlying environment requires decomposing the environment into its computationally elemental fragments. We suggest how to find fragmentations using unsupervised, mixtures model, learning methods on data derived from optimal value functions for multiple tasks, and show that these fragmentations are in accord with observable structures in the environments. Further, we present evidence that such fragments can be of use in a practical reinforcement learning context, by facilitating online, actor-critic learning of multiple goals MDPs.

Original languageEnglish (US)
Pages (from-to)325-346
Number of pages22
JournalMachine Learning
Volume49
Issue number2-3
DOIs
StatePublished - Nov 2002

Keywords

  • Density estimation
  • Dynamic programming
  • Mixture models
  • Reinforcement learning
  • Unsupervised learning
  • Value functions

ASJC Scopus subject areas

  • Software
  • Artificial Intelligence

Fingerprint Dive into the research topics of 'Structure in the space of value functions'. Together they form a unique fingerprint.

Cite this