Flinders University Flinders Academic Commons

Flinders Academic Commons >
Flinders Digital Archive >
Science and Engineering >
Computer Science, Engineering and Mathematics  >
Computer Science, Engineering and Mathematics - Collected Works >

Please use this identifier to cite or link to this item: http://hdl.handle.net/2328/26286

Title: Gain/variability tradeoffs in undiscounted Markov Decision Processes
Authors: Filar, Jerzy A
Lee, Huey-Miin
Keywords: Mathematics
Markov Decision Processes
Issue Date: 1985
Publisher: Institute of Electrical and Electronic Engineers
Citation: Filar, J.A. and Lee, H.-M., 1985. Gain/variability tradeoffs in undiscounted Markov Decision Processes. Proceedings of the 24th IEEE Conference on Decision and Control, vol. 24(1), 1106-1112.
Abstract: We consider a finite state/action Markov Decision Process over the infinite time horizon, and with the limiting average reward criterion. However, we are interested not only in maximizing the above reward criterion but also in minimizing "the variability" of the stream of rewards. The latter notion is formalized in two alternative ways: one in terms of measuring absolute deviations from the "optimal" reward, and the other in terms of a "long-run variance" of a policy. In both cases we formulate a bi-objective optimization problem and show that efficient (i.e., "nondominated") deterministic stationary policies exist and can be computed by finite algorithms. In addition, in the former case we give an algorithm for computing a finite set of "critical efficient policies" which in a sense constitutes one complete set of reasonable responses by a decision-maker sensitive to the variability of rewards. However, the analysis of this case is intended primarily as a "sensitivity analysis" technique rather than a complete theoretical treatment of the gain/variability tradeoffs.
URI: http://hdl.handle.net/2328/26286
Appears in Collections:Computer Science, Engineering and Mathematics - Collected Works

Files in This Item:

File Description SizeFormat
Filar Gain variability.pdf681.42 kBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.


Valid XHTML 1.0! DSpace Software Copyright © 2002-2010  Duraspace - Feedback