A Lagrangian decomposition scheme for choice-based optimization

• Published in: Computers & operations research 2022-12, Vol.148, p.105985, Article 105985
• Main Authors: Pacheco Paneque, Meritxell, Gendron, Bernard, Sharif Azadeh, Shadi, Bierlaire, Michel
• Format: Article
• Language: English
• Subjects: Choice-based optimization; Combinatorial optimization; Lagrangian decomposition; Random utility models; Scenario decomposition; Scenario grouping
• ISSN: 0305-0548; 1873-765X
• DOI: 10.1016/j.cor.2022.105985

Choice-based optimization problems are the family of optimization problems that incorporate the stochasticity of individual preferences according to discrete choice models to make planning decisions. This integration brings non-convexity and nonlinearity to the associated mathematical formulations. Previously, the authors have tackled these issues by introducing a simulation-based approximation of the choice model with the aim of linearizing it. Nevertheless, already existing exact methods and state-of-the-art commercial solvers fail to solve relevant instances. In this paper, we propose a novel Lagrangian decomposition method inspired by scenario decomposition and scenario grouping in the stochastic programming framework for the purpose of solving choice-based optimization problems. In addition, we develop a tailored algorithm to generate feasible solutions to the original problem from the solution of the Lagrangian subproblem. Hence, at each iteration of the subgradient method, which is used to solve the Lagrangian dual, we provide both an upper and a lower bound to the original problem. This enables the calculation of the duality gap to assess the quality of the generated solutions. Computational results show that the decomposition method provides solutions with optimality gaps below 0.5% and restricted duality gaps within low computational times. We also show that scenario grouping leads to high-quality feasible solutions and lower duality gaps. •Lagrangian decomposition method for a general choice-based optimization problem.•Decomposition by scenario groups to improve the obtained upper bound.•Exploring feasible solutions via an algorithm that solves a restricted formulation.