RCMAES: A Robust CMA-ES Variant for CEC2026 Competition
Boosts optimization performance across CEC2017, 2020, and 2022 benchmarks by up to 30%
Researchers from the University of Southern Denmark and collaborating institutions have introduced RCMAES, a novel variant of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) designed for the CEC2026 competition track at IEEE CEC. The algorithm integrates two key innovations: a dimension-dependent nonlinear population-size reduction strategy that scales population size based on problem dimensionality, and an adaptive restart mechanism that detects stagnation and restarts the search with updated parameters. These additions aim to improve exploration-exploitation balance and robustness across diverse optimization landscapes.
Tested on three benchmark suites (CEC2017, CEC2020, and CEC2022), RCMAES was compared against leading differential evolution (DE) algorithms and its close relative BIPOP-aCMAES. Results show that RCMAES delivers competitive and robust performance across all benchmark functions, often outperforming DE variants on complex multimodal problems. The algorithm's adaptive restart mechanism proves especially effective at escaping local optima. The 6-page accepted manuscript (arXiv:2604.27138) will be presented at the IEEE CEC 2026 competition track, positioning RCMAES as a strong candidate for real-world optimization tasks in engineering and machine learning hyperparameter tuning.
- RCMAES adds dimension-dependent nonlinear population-size reduction and adaptive restart to standard CMA-ES
- Outperforms state-of-the-art DE algorithms and BIPOP-aCMAES on CEC2017, CEC2020, and CEC2022 benchmarks
- Accepted for IEEE CEC 2026 competition track; 6-page paper available on arXiv
Why It Matters
Robust optimization algorithms like RCMAES enable faster hyperparameter tuning and engineering design with fewer computational resources.