Learning-aided metaheuristic framework using radial basis function modeling

doi:10.1016/j.asoc.2026.116181

Learning-aided metaheuristic framework using radial basis function modeling

2026-08-01

0

PRE

AI

DOI：10.1016/j.asoc.2026.116181

Original

Original request for help

Save

Abstract

En 中文

Most Metaheuristic Algorithms (MAs) lack explicit learning capabilities. Although they generate a large number of candidate solutions during the iterative search, they typically rely only on the information contained in the current population to guide the search process. While this design simplifies implementation, the resulting decisions are largely driven by random strategies, preventing the algorithm from inferring the underlying topology of the search landscape. This blind behavior increases the risk of premature convergence, reduces population diversity, and often leads to the search for unpromising regions. Although several learning-enhanced and surrogate-assisted MAs have been proposed, these approaches generally focus on approximating the objective function alone. Consequently, they remain limited in their ability to identify directions that previously led to improvements, detect promising regions of the search space, and recognize poorly performing areas that should be avoided. This paper proposes a novel learning-aided method that reuses candidate solutions from previous iterations to guide the search strategy without approximating the objective function. The proposed approach extracts information using a Radial Basis Function (RBF) to distinguish promising and poor-quality search regions. To promote effective and diverse initial coverage, the population is initialized using the MetropolisHastings (MH) algorithm. In each generation, a subset of solutions is used to train the RBF. The learned model is then exploited to generate new candidate solutions in locally promising regions of the search space. The use of the RBF and MH mechanisms enhances the convergence to complex, high-dimensional, and multimodal search spaces, avoiding revisiting poor-quality regions. The performance of the proposed methodology is compared to several state-of-the-art MAs considering a set of well-known functions evaluated in low and high dimensional search spaces. The results demonstrate that the proposed learning-aided strategy achieves an effective balance between accuracy and computational efficiency over its counterparts.

Keywords:

Radial-basis function

Metaheuristics

Extracting Knowledge

Non-linear optimization

Machine learning

Learning-aided metaheuristic framework using radial basis function modeling

Abstract

Journal

Organization

Cited Papers