Time consistent expected mean-variance in multistage stochastic quadratic optimization: a model and a matheuristic

In this paper, we present a multistage time consistent Expected Conditional Risk Measure for minimizing a linear combination of the expected mean and the expected variance, so-called Expected Mean-Variance. The model is formulated as a multistage stochastic mixed-integer quadratic programming proble...

ver descrição completa

Detalhes bibliográficos
Autores: Aldasoro Marcellan, Unai, Merino Maestre, María, Pérez Sainz de Rozas, Gloria
Formato: artículo
Fecha de publicación:2019
País:España
Recursos:Universidad del País Vasco
Repositorio:Addi. Archivo Digital para la Docencia y la Investigación
OAI Identifier:oai:addi.ehu.eus:10810/64954
Acesso em linha:http://hdl.handle.net/10810/64954
Access Level:acceso abierto
Palavra-chave:branch-and-fix coordination
expected conditional risk measures
matheuristic algorithms
McCormick relaxation
multistage stochastic optimization
quadratic mixed 0–1 programming
production planning
time consistent risk aversion
Descrição
Resumo:In this paper, we present a multistage time consistent Expected Conditional Risk Measure for minimizing a linear combination of the expected mean and the expected variance, so-called Expected Mean-Variance. The model is formulated as a multistage stochastic mixed-integer quadratic programming problem combining risk-sensitive cost and scenario analysis approaches. The proposed problem is solved by a matheuristic based on the Branch-and-Fix Coordination method. The multistage scenario cluster primal decomposition framework is extended to deal with large-scale quadratic optimization by means of stage-wise reformulation techniques. A specific case study in risk-sensitive production planning is used to illustrate that a remarkable decrease in the expected variance (risk cost) is obtained. A competitive behavior on the part of our methodology in terms of solution quality and computation time is shown when comparing with plain use of CPLEX in 150 benchmark instances, ranging up to 711,845 constraints and 193,000 binary variables.