On the Risk-Sensitive Optimality Criteria for Markov Decision Processes.
Abstract
Discrete dynamic programming models with an exponential utility function are studied with respect to the asymptotic behavior of the dynamic programming recursion for the expected utility. Preliminary results on maximizing the asymptotic growth of the expected utility in the class of stationary policies are presented. Under the condition that there exists a stationary 'optimal' policy with an irreducible, aperiodic transition probability matrix, some nice limiting properties for the maximum expected utilities are established. Moreover, it is shown how to generate a monotonic sequence of lower and upper bounds on the maximum growth rate of the expected utility. Under certain additional assumptions it is possible to extend the obtained results to Markov decision processes with a denumerable state space.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 15, 1975
- Accession Number
- ADA016234
Entities
People
- Karel Sladky
Organizations
- Stanford University