Some Results in Dynamic Programming
Abstract
The first part of the report discusses a dynamic programming model in which all rewards obtained by the decision maker are assumed nonnegative. The decision maker's objective is to successively choose actions so as to maximize his expected reward earned over an infinite time span. It follows from known results that the decision maker's choice need only depend upon the outcome of a randomization that depends on the model only through the state of the model and the time when the choice is made. It is shown by counterexample that this is basically the smallest class of decision rules that need be considered. Conditions under which a stationary policy is optimal are also presented. The second part of the report discusses the same model under a new criteria, namely, the average cost incurred per unit time. An example is presented in which there does not exist an epsilon-optimal randomized stationary policy.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1971
- Accession Number
- AD0722347
Entities
People
- Sheldon M. Ross
Organizations
- University of California, Berkeley