A methodology for multicriteria stochastic anticipatory optimization

Abstract

This research project aims to design new sequential decision-making systems, operating in uncertain environments under multiple conflicting optimization criteria. It is assumed that the dynamics of the system under control (in discrete time and over a finite horizon) is linear and that the exogenous uncertainty can be estimated by parametric probabilistic models. In this context, four challenges are covered, namely: (1) the learning of probabilistic models capable of measuring the influence of the decisions implemented on the future operating costs; (2) the determination of stochastic policies capable of modeling the decision maker; (3) the determination of the risks of violating the problem constraints; and (4) the incorporation of partial preferences in the decision making process. It is emphasized that research activity on the treatment of multiple conflicting criteria and the incorporation of chance-constraints is scarce, considering the literature of anticipatory meta-heuristics and approximate dynamic programming. As its main contribution, this project proposes a new methodology as well as tools to allow for the effective synthesis of anticipatory multicriteria decision-making systems. The methodology will be investigated over a broad class of problems, ranging from vendor managed inventory-routing problems; optimization of financial and product portfolios, and the control of public transport systems operating in real time.