.png){kind=icon}
Return
Mega-scale evolutionary optimization: The Descent Direction Distribution algorithm
.png){kind=icon}
PRE
AI
DOI：10.1016/j.swevo.2026.102454.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Large-scale black-box optimization has traditionally addressed problems with up to 103-104 variables, where many evolutionary algorithms exhibit superlinear time or memory complexity. Emerging mega-scale scenarios, however, demand derivative-free optimization with up to 106 variables under strict computational constraints. We introduce the Descent Direction Distribution (DDD) algorithm, an Estimation of Distribution Algorithm designed for mega-scale continuous optimization with (9(n) time and memory complexity and a constant population size. DDD combines a diagonal exploration model with a one-dimensional projection model that estimates a descent direction via least-squares regression, preserving strict linear scalability while retaining directional search capability. In addition, DDD incorporates an adaptive stagnation-detection mechanism that mitigates premature convergence without requiring warm-up phases. An extensive experimental study including separable and non-separable benchmark functions from 103 to 106 variables compares DDD against LMMAES, LSHADE, and NLSHADE-RPS. Results show stable scalability and competitive, often superior, performance in ultra-high-dimensional regimes under limited evaluation budgets. These findings demonstrate that strictly linear-complexity evolutionary optimization in the mega-scale black-box setting is both feasible and practically effective.
Keywords:
Evolutionary algorithm
Directional search
Variance direction
Ultra-high-dimensional black-box optimization
Large scale optimization
Journal
IF:
8.5
Papers:
2.2K
Citations:
1.0W
Organization
No organization information available
Cited Papers
No cited papers available