Science China Information Sciences General Computer Science
A multiregional lagrangian-dual-based algorithm for dynamic constrained multiobjective optimization
Abstract
In dynamic constrained multiobjective optimization problems, the Pareto fronts (PFs) take on various appearances over time, influenced by changes in the unconstrained Pareto fronts and feasible regions. Aimed at accurately and rapidly tracking the changing PF, this paper proposes a multiregional Lagrangian-dual-based dynamic constrained multiobjective optimization algorithm (MLDCMOA). When a change occurs, a multiregional prediction strategy, involving multiregional division, population prediction, and population correction mechanisms, is triggered to estimate the new Pareto optimal set. To handle constraints, a Lagrangian-dual-factor-based technique is tailored to leverage feasibility, convergence, and diversity. It converts the original objectives of individuals and assigns their priorities, which are further embedded with the population selection operator to increase the selection pressure toward the PF in the improved optimizer. Moreover, a hybrid offspring generation approach, derived from the proposed M -objective search strategy, is designed to enhance the spread of the solutions and facilitate the capability of breaking the constrained limitation. To verify its effectiveness, MLDCMOA is compared with four state-of-the-art algorithms on two newly proposed test suites and one practical case. Experimental results demonstrate the superiority of MLDCMOA in providing a competitive forecast and accelerating convergence in a dynamic and constrained environment.