Ensemble Machine Learning Helps Evolutionary Algorithms Track Changing Optimization Targets

Dynamic multi-objective optimization problems, in which the objectives themselves shift over time, arise in many decision-making scenarios. A study published in Cluster Computing addresses how such algorithms respond to these shifts with EnsembleML, a prediction mechanism that uses ensemble machine learning to reinitialize the population after each detected change, and reports that it outperformed eight existing response strategies across a standard benchmark. When EnsembleML detects an environmental change, it first builds a training dataset from historical Pareto fronts, the sets of trade-off solutions that previous environments produced.

It then trains four complementary regressors: Linear Regression, Support Vector Regression, XGBoost, and Random Forest. These models predict the key points of the new Pareto front, and Gaussian perturbation and mutation are applied to generate five sub-populations around those predictions. The sub-populations are then merged to form the initial population for the next environment.

The researchers evaluated the approach on the CEC 2018 test suite, which comprises 14 dynamic multi-objective optimization problems, and compared EnsembleML against eight state-of-the-art response mechanisms. Performance was measured with two standard indicators: the Mean Inverted Generational Distance (MIGD) and the Mean HyperVolume Difference (MHVD). Under severe environmental changes, EnsembleML achieved the best MIGD values in 6 of the 14 problems and the best MHVD values in 10 problems.

Under mild changes, it recorded the best MIGD in 4 problems and the best MHVD in 5 problems. The counts show that the mechanism led on many, but not all, of the benchmark problems and metrics. The authors state that these results demonstrate the effectiveness of the proposed ensemble approach.

The study's contribution lies in the environmental response mechanism, the component of a dynamic multi-objective evolutionary algorithm that reinitializes the population when conditions shift. Rather than relying on a single forecasting model, EnsembleML combines four regressors with different characteristics and uses their predictions to seed the population for the next environment through Gaussian perturbation and mutation. Because the evaluation is confined to the CEC 2018 benchmark problems, the findings are measured observations on simulated test cases rather than evidence of performance in specific applied settings.

The comparison covers the 14 benchmark problems and the eight response mechanisms included in the study, and the reported advantage varies by problem, metric, and severity of change. Further work would be needed to establish how the mechanism behaves on real-world dynamic optimization tasks. Within the scope of the benchmark study, however, the results indicate that an ensemble of complementary regressors, trained on historical Pareto fronts, can serve as an effective response mechanism for a dynamic multi-objective evolutionary algorithm when the environment changes.

Subject of Research: Ensemble machine learning-based environmental response mechanisms for dynamic multi-objective evolutionary algorithms

Article Title: An ensemble machine learning-based dynamic multi-objective evolutionary algorithm

Article References: Shao, K., Wang, Y., Wang, B., Sang, Y., Chen, M., Pi, X., & Li, H. (2026). An ensemble machine learning-based dynamic multi-objective evolutionary algorithm. Cluster Computing, 29(16), 896. https://doi.org/10.1007/s10586-026-06729-y

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06729-y

Keywords: machine learning, evolutionary computation, dynamic multi-objective optimization, ensemble learning, XGBoost, random forest, support vector regression, Pareto front prediction, CEC 2018 benchmark, Cluster Computing, population reinitialization, environmental response mechanism