On Threshold Properties of the Optimal Policy for POMDPs on Partially Ordered Spaces
Erik Miehling

Citation
Erik Miehling. "On Threshold Properties of the Optimal Policy for POMDPs on Partially Ordered Spaces". Talk or presentation, 23, August, 2017.

Abstract
Partially observable Markov decision processes (POMDPs) arise frequently in real-world problems. Unfortunately, problems in this setting are typically large, usually precluding one from finding an optimal policy. In order to alleviate the complexity of this search, we investigate if there exist conditions that allow one to say that the optimal policy has any special structure. Specifically, we study a specific class of POMDPs (problems with a binary action space) and obtain conditions on the state and observation processes such that the resulting policy is monotone in the belief state. We consider general partially ordered state and observation spaces, requiring us to extend the existing theory. Applications in security are discussed.

Electronic downloads

Internal. This publication has been marked by the author for FORCES-only distribution, so electronic downloads are not available without logging in.

Citation formats

• HTML

  Erik Miehling. <a
  href="http://www.cps-forces.org/pubs/277.html"
  ><i>On Threshold Properties of the Optimal Policy
  for POMDPs on Partially Ordered Spaces</i></a>,
  Talk or presentation,  23, August, 2017.

• Plain text

  Erik Miehling. "On Threshold Properties of the Optimal
  Policy for POMDPs on Partially Ordered Spaces". Talk or
  presentation,  23, August, 2017.

• BibTeX

  @presentation{Miehling17_OnThresholdPropertiesOfOptimalPolicyForPOMDPsOnPartially,
      author = {Erik Miehling},
      title = {On Threshold Properties of the Optimal Policy for
                POMDPs on Partially Ordered Spaces},
      day = {23},
      month = {August},
      year = {2017},
      abstract = {Partially observable Markov decision processes
                (POMDPs) arise frequently in real-world problems.
                Unfortunately, problems in this setting are
                typically large, usually precluding one from
                finding an optimal policy. In order to alleviate
                the complexity of this search, we investigate if
                there exist conditions that allow one to say that
                the optimal policy has any special structure.
                Specifically, we study a specific class of POMDPs
                (problems with a binary action space) and obtain
                conditions on the state and observation processes
                such that the resulting policy is monotone in the
                belief state. We consider general partially
                ordered state and observation spaces, requiring us
                to extend the existing theory. Applications in
                security are discussed.},
      URL = {http://cps-forces.org/pubs/277.html}
  }
  

Posted by Carolyn Winter on 24 Aug 2017.
Groups: forces
For additional information, see the Publications FAQ or contact webmaster at cps-forces org.