Denumerable State Markov Decision Processes with Unbounded Costs.
Abstract
The report establishes sufficient conditions for both the existence of stationary optimal policies and the optimality of stationary policies in Markov decision processes with unbounded costs. The otpimization criteria considered are minimum expected discounted cost over an infinite horizon and minimum expected average cost per unit time. Sufficient conditions that one may frequently establish in applications are given for the existence of a stationary optimal policy for both optimization criteria. It is also shown that for both optimization criteria optimal stationary policies are associated with the solution of the usual functional equations that arise in Markov decision processes with bounded costs. With unbounded costs, however, one must place additional constraints on these solutions to assure that the implied stationary policy is optimal. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 06, 1973
- Accession Number
- AD0771432
Entities
People
- Frank C. Reed
Organizations
- Stanford University