Supervised learning-guided individual-level operator selection in differential evolution for constrained optimization

doi:10.1016/j.knosys.2026.116813

Supervised learning-guided individual-level operator selection in differential evolution for constrained optimization

2026-08-29

0

OA

AI

DOI：10.1016/j.knosys.2026.116813

Original

Original request for help

Save

Abstract

En 中文

Adaptive operator selection in differential evolution has been widely addressed through reinforcement learning and heuristic credit-assignment methods. However, no prior work has applied supervised machine learning to predict and assign mutation operators at the individual level for constrained optimization problems. We present SL-MODE, a supervised learning-guided multi-operator DE framework for constrained optimization. Unlike reinforcement learning-based selectors, which require reward engineering, and policy networks, SL-MODE trains a lightweight supervised classifier online, directly from population feedback, without offline pre-training. At each generation, the classifier predicts the most promising mutation operator for each individual separately, conditioned on the current search state. Training is guided by a continuous improvement index that jointly encodes objective progress and constraint-violation reduction, giving the model an explicit feasibility-aware learning signal. The framework is further enhanced by two supporting algorithmic components: a hybrid initialization technique that aims to balance global diversity with adaptive intensification around promising regions, and an adaptive equality-constraint relaxation. The results on the CEC2006, CEC2010 constrained benchmark suites and 33 real-world problems from CEC2020 demonstrate that SL-MODE consistently achieves high feasibility rates, high-quality solutions, and competitive performance with statistically significant advantages over several competing algorithms. On the CEC2006 problems, sensitivity analyses conducted confirm robustness across a wide range of algorithmic parameters.

Keywords:

Evolutionary algorithms

Supervised learning

Constrained optimization

Population initialization

Supervised learning-guided individual-level operator selection in differential evolution for constrained optimization

Abstract

Journal

Organization

Cited Papers