A two-stage ensemble evolutionary algorithm for constrained multi-objective optimization

In constrained multi-objective evolutionary algorithms (CMOEAs), selecting appropriate constraint-handling techniques (CHTs) is challenging without prior knowledge of the problem’s constraint severity or feasible region distribution. Ensemble frameworks that integrate multiple CHTs with distinct pop...

Full description

Saved in:
Bibliographic Details
Published in:Swarm and evolutionary computation Vol. 99; p. 102213
Main Authors: Modampuri, Sri Srinivasa Raju, Fan, Jiahao, Sun, Yanan
Format: Journal Article
Language:English
Published: Elsevier B.V 01.12.2025
Subjects:
Constrained multi-objective optimization problem
Constraint handling technique
Ensemble framework
Evolutionary algorithm
Multi-armed bandit
Ensemble framework
Constrained multi-objective optimization problem
Evolutionary algorithm
Multi-armed bandit
Constraint handling technique
ISSN:2210-6502
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In constrained multi-objective evolutionary algorithms (CMOEAs), selecting appropriate constraint-handling techniques (CHTs) is challenging without prior knowledge of the problem’s constraint severity or feasible region distribution. Ensemble frameworks that integrate multiple CHTs with distinct populations offer a promising solution but face issues like redundant evaluations and poor exploration–exploitation balance. To address these limitations, we propose a two-stage ensemble-based CMOEA (CMOEA-TENS) that dynamically prioritizes suitable CHTs based on problem characteristics. Specifically, in the first stage, a population dedicated to explore the unconstrained search space drives the evolutionary process, while remaining populations co-evolve by leveraging solutions identified by the exploratory population. In the second stage, an ensemble of distinct populations drives the evolutionary process, each co-evolving with a different CHT focused on feasibility, diversity, or convergence to exploit the feasible regions effectively. Furthermore, we introduce a novel Multi-Armed Bandit (MAB)-based decision-making strategy that, unlike existing static or random selection approaches, adaptively learns and selects the most suitable CHT-based population to drive the evolutionary process based on real-time performance feedback. This dynamic strategy explicitly reduces redundant functional evaluations and ensures better management of exploration–exploitation trade-offs. CMOEA-TENS was evaluated against eleven state-of-the-art algorithms across six popular test suites, encompassing 57 test instances and six real-world problems. The empirical results demonstrate that CMOEA-TENS effectively balances exploration and exploitation while avoiding redundant evaluations by dynamically selecting the most suitable CHT-based population to drive the evolutionary process. Additionally, an ablation study further validates the effectiveness of the designed MAB strategy. •Proposed a two-stage ensemble CMOEA for solving constrained multi-objective problems.•Developed a dynamic two-stage strategy balancing exploration and exploitation effectively.•Designed an MAB strategy to select suitable population for offspring generation.•Validated the proposed algorithm on six test suites and six real-world optimization tasks.
ISSN:2210-6502
DOI:10.1016/j.swevo.2025.102213