Bayesian Quadratic Reinforcement Learning

Philipp Hennig, David L. Stern, Thore GraepelMPG.PuRe 2009

Abstract

The idea of using "optimistic" evaluations to guide exploration has been around for a long time, but has received renewed interest recently thanks to the development of bound-based guarantees [Kolter and Ng, 2009].Asmuth et al. [2009] have recently presented a way to construct optimistic evaluations by merging sampled MDPs.Their BOSS algorithm is polynomial in the size of the state-action-space of the merged MDP, and thus a multiple of the size of the actual state space.
BibTeX
@inproceedings{hennig2009bayesian,
  author = {Philipp Hennig and David L. Stern and Thore Graepel},
  title = {{Bayesian Quadratic Reinforcement Learning}},
  booktitle = {MPG.PuRe},
  year = {2009},
}

Loading related content...

Figure