Κατεύθυνση Στατιστική και Επιχειρησιακή Έρευνα
Library of the School of Science

2014-01-29

2014

Δημητρακόπουλος Ιωάννης

Απόστολος Μπουρνέτας Καθηγητής ΕΚΠΑ (επιβλέπων), Αντώνιος Οικονόμου Αναπλ. Καθηγητής ΕΚΠΑ, Κωνσταντίνος Μηλολιδάκης Αναπλ. Καθηγητής ΕΚΠΑ

Δυναμικός προγραμματισμός με χρήση γεγονότων

Greek

Event-based dynamic programming

In this thesis, we present Event-based dynamic programming, an alternative
method for analyzing the value function of Markov decision processes (MDP's)
and deriving the optimal policy for a variety of control problems. This method
analyzes in a systematic manner the monotonicityproperties of the value
function simplifying their proofs, in contrast to the use of classic induction
in the value function which refers to successive approximations. The main idea
of the method is to rewrite the value function as the composition of event
operators which refer to structural elements of the value function and not
necessarily to events that affects the development of the process. In this
shape, we derive monotonicityproperties of the value function studying the
properties that event operators satisfy for any real function defined in the
state space of the problem. In this thesis, we present the theoretical
background of this method and enlist several event operators along with their
monotonicity properties, referring to Koole(2007)who has systematically
developed this method. Finally, we formulate and examine the joint admission
and service control model of Dimitrakopoulos and Burnetas (2011) under this
framework.

Event-based dynamic programming, Queueing control, Markov decision processes, Monotonicity in queueing, Event operators

No

0

No

25

77

File access is restricted.

https://pergamos.lib.uoa.gr/uoa/dl/object/1317045