22 January 2019
Abstract
Many multi-stage decisions under uncertainty can be structured as influence diagrams which consist of decision, chance and value nodes as well as arcs representing dependencies between these nodes. The decision strategy which maximizes the decision maker's (DM) expected utility is typically determined either by carrying local transformations (such as arc reversals and node removals) or by formulating the equivalent decision tree which is then solved with dynamic programming.
This presentation describes an approach called Decision Programming (DP) in which such decision models are solved by transforming them into equivalent linear programming problems. In the context of project portfolio optimization, Decision Programming can be viewed as an extension of Contingent Portfolio Programming (CPP) to problems in which project decisions can impact the scenario tree probabilities. More generally, Decision Programming offers enhanced modelling possibilities in that (i) there is no need for the usual ‘no forgetting’ assumption in that earlier decisions need not be known when making later ones, (ii) both deterministic and chance constraints can be handled, and (iii) different risk preferences can be accounted for by modifying the objective function. We present illustrative examples and provide evidence on computational performance.
PUBLICATIONS
Research projects
International Institute for Applied Systems Analysis (IIASA)
Schlossplatz 1, A-2361 Laxenburg, Austria
Phone: (+43 2236) 807 0 Fax:(+43 2236) 71 313