Parallel strategies for Direct Multisearch

作     者：Tavares, S. Bras, C. P. Custodio, A. L. Duarte, V Medeiros, P. 

作者机构：FCT NOVA Campus Caparica P-2829516 Caparica Portugal FCT NOVA CMA Dept Math Campus Caparica P-2829516 Caparica Portugal FCT NOVA NOVA LINCS Dept Comp Sci Campus Caparica P-2829516 Caparica Portugal 

出 版 物：《NUMERICAL ALGORITHMS》 (数值算法)

年 卷 期：2023年第92卷第3期

页      面：1757-1788页

核心收录：

学科分类：07[理学] 070104[理学-应用数学] 0701[理学-数学] 

基　　金：FCT -Fundacao para a Ciencia e a Tecnologia [PTDC/MAT-APL/28400/2017, UIDB/00297/2020, UIDP/00297/2020] Fundação para a Ciência e a Tecnologia [PTDC/MAT-APL/28400/2017, UIDP/00297/2020] Funding Source: FCT 

主　　题：Multiobjective optimization Derivative-free optimization Direct search methods Parallel algorithms 

摘      要：Direct multisearch (DMS) is a derivative-free optimization class of algorithms, suited for computing approximations to the complete Pareto front of a given multiobjective optimization problem. In DMS class, constraints are addressed with an extreme barrier approach, only evaluating feasible points. It has a well-supported convergence analysis and simple implementations present a good numerical performance, both in academic test sets and in real applications. Recently, this numerical performance was improved with the definition of a search step based on the minimization of quadratic polynomial models, corresponding to the algorithm BoostDMS. In this work, we propose and numerically evaluate strategies to improve the performance of BoostDMS, mainly through parallelization applied to the search and to the poll steps. The final parallelized version not only considerably decreases the computational time required for solving a multiobjective optimization problem, but also increases the quality of the computed approximation to the Pareto front. Extensive numerical results will be reported in an academic test set and in a chemical engineering application.