Risk-averse control of continuous-time markov chains

Darinka Dentcheva, Andrzej Ruszczynski

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Scopus citations

Abstract

We develop an approach to time-consistent risk control of continuous-time processes in Markov systems with finite state space. Our analysis is based on time-consistent risk evaluation in continuous time, which uses the dual representation of dynamic coherent risk measures. Introducing a suitable differentiability concept for multivalued mappings and a concept of strong time consistency, we approximate the risk multikernels of the Markov system via set-valued derivatives, which gives rise to the concept of a risk multigenerator of a Markov process. We focus on risk-averse Markov decision problems with cost evaluated at the terminal state. We derive a system of ordinary differential equations for the risk evaluation of a given policy, which generalize the classical backward Kolmogorov equations for Markov processes. Furthermore, we establish optimality conditions in form of differential equation involving the support function of the risk multigenerator and we identify conditions for the existence of an optimal Markov policy.

Original languageEnglish
Title of host publicationProceedings of the SIAM Conference on Control and Its Applications, CT 2017
EditorsWei Kang, Jean-Pierre Barbot, Qing Zhang, Ruihua Liu
Pages78-85
Number of pages8
ISBN (Electronic)9781611975000
DOIs
StatePublished - 2017
Event2017 SIAM Conference on Control and Its Applications, CT 2017 - Pittsburgh, United States
Duration: 10 Jul 201712 Jul 2017

Publication series

NameProceedings of the SIAM Conference on Control and Its Applications, CT 2017

Conference

Conference2017 SIAM Conference on Control and Its Applications, CT 2017
Country/TerritoryUnited States
CityPittsburgh
Period10/07/1712/07/17

Fingerprint

Dive into the research topics of 'Risk-averse control of continuous-time markov chains'. Together they form a unique fingerprint.

Cite this