ARRDE algorithm delivers robust optimization across 5 CEC benchmark suites

A new paper on arXiv introduces the Adaptive Restart–Refine Differential Evolution (ARRDE) algorithm, aiming to solve a persistent problem in evolutionary computation: most optimization algorithms perform well only within narrow problem regimes and degrade when applied to others. ARRDE explicitly targets robustness across heterogeneous optimization scenarios—varying dimensionality, landscape structures, and evaluation budgets—making it attractive for practitioners who need a one-size-fits-all solver.

ARRDE's three core innovations drive its stability. First, an adaptive restart–refine strategy periodically resets or refines the population to escape local optima. Second, a nonlinear population-size reduction schedule scales down the population based on problem dimensionality, balancing exploration and exploitation. Third, a budget-aware initialization rule sets the initial population size based on the available evaluation budget, ensuring efficiency under limited resources. These mechanisms work together to maintain performance without manual tuning across different problems.

The authors conduct an unusually comprehensive evaluation on five benchmark suites: CEC2011, CEC2017, CEC2019, CEC2020, and CEC2022. These suites differ in dimensionality (from 10 to 100), landscape properties (unimodal, multimodal, hybrid, composite), and evaluation budgets. Since each suite uses different official metrics, the authors also propose a bounded accuracy-based scoring metric derived from relative error, enabling fair cross-suite comparison. ARRDE consistently ranks among the top performers, demonstrating one of the most stable aggregate profiles reported for a DE variant.