Department of Mathematical Sciences
Division of Statistics and Probability
SEMINAR
A New Approach for Exploring Exhaustively the Pareto Set of Multi Criteria
Markov Decision Processes with Finite State and Action Sets.
Gianluca Dorini, University of Exeter
Wednesday, 9th February 2005, 2pm
The Whittaker Room (211)
Abstract:
Constrained Linear Programming (CLP) and Weighted Dynamic Programming (WDP)
are the methods most commonly used for solving finite Multi Criteria Markov
Decision Processes (MCMDP). First a set of constraint values (CLP) or
weights (WDP) is chosen to reduce the Multi Objective Problem (MOP) to a
Single Objective Problem (SOP). Then the associated SOP is solved and a
single Pareto optimal Policy is determined. In order to explore the Pareto
set, this two-step procedure must be repeatedly iterated by opportunely
varying constraints and weights.
This seminar focuses on a new technique to explore the Pareto Set of Multi
Criteria Markov Decision Processes. It combines the knowledge of the
structural properties of the Pareto set of MCMDPs with concepts and tools
drawn from Computational Geometry to effectively exploit it. As opposed to
the aforementioned approaches, each single run of the technique proposed
generates and infinite number of Pareto optimal policies
Following the talk, tea and biscuits will be available in Room 304
ALL WELCOME
********************************************************
Ingrid Harper
University of Liverpool
Department of Mathematical Sciences
Division of Statistics and Probability
Peach Street
Liverpool L69 7ZL
Tel: 0151 794 4751
Fax: 0151 794 4754
********************************************************
|