Cogent Engineering

Development of Novel Hybrid Methods based on Machine Learning and Metaheuristics

In recent years, considerable excitement has surrounded around research in machine learning, which has noticeably reduced the number of publications in some traditional areas, such as metaheuristics. New methods dominated by neural networks have surpassed traditional methods in many fields, and a significant number of researchers have switched to this new discipline. However, it may be interesting to examine how the advantages of the two approaches can be combined to find novel, even more efficient methods. This combination could be beneficial in both directions (and there are already some examples of such approaches): there are metaheuristic implementations that use neural networks in their operation, but the reverse is also true, as numerous publications report that some form of metaheuristics was used in the training of neural networks (either to build the appropriate architecture or to adjust the weights themselves).

There are already preliminary results in this area, showing that the potential of machine learning can be incorporated into existing metaheuristics in several ways. The most common example is the so-called metamodeling, where an attempt is made to replace computationally expensive fitness calculations with a faster, less accurate machine-learning approach. Another common approach is attempting to estimate the random starting position of a given metaheuristic using a machine learning model that provides a roughly acceptable approximation for the final result. Another hybrid approach is to determine the stopping condition of metaheuristics using a machine learning model, which indicates whether it is worth continuing the search or better to start a new one. Numerous publications have appeared in these domains, and relevant review articles are also available. However, these studies are mostly very limited and usually focus on solving a specific practical task rather than proposing novel, generally usable methods.

It would be worthwhile to examine different general-purpose hybrid implementations: local search methods in which the step size and direction in a given iteration are not determined randomly, but instead based on an intelligent model. An interesting direction could be to examine and tune specific metaheuristics. For example, in tabu search, instead of storing the tabu list, a machine learning model could memorize and monitor whether a given location has already been visited. Another example can be a modified simulated annealing algorithm, where a machine learning model could be used to determine whether a step in the wrong direction is allowed, rather than the usual temperature-based method.

This Articles Collection aims to bring together current developments in the field of machine learning-metaheuristics hybrids. The Articles Collection focuses on combinations of various metaheuristics (hill-climbing, gradient-based, evolution-based, population-based approaches, etc.) and various machine learning models (deep neural networks, decision trees, SVMs, clustering, etc.). Articles may relate to hybrid versions of existing, well-known methods or to the design of entirely new solutions. In terms of the generalizability of the results, both the design of novel, generally applicable algorithms and the experience gained from solving specific practical tasks may be of interest.

Keywords: Metaheuristics, Machine learning, Hybrid methods, Optimization, Deep learning

­­All manuscripts submitted to this Article Collection will undergo a full peer-review; the Guest Advisor for this Collection will not be handling the manuscripts (unless they are an Editorial Board member).

Please review the journal scope and author submission instructions prior to submitting a manuscript.

The deadline for submitting manuscripts is 31 July 2026.

Please contact Hang Ke at [email protected] with any queries and discount codes regarding this Article Collection.

Please be sure to select the "Development of Novel Hybrid Methods based on Machine Learning and Metaheuristics" from the drop-down menu in the submission system.

Sándor Szénási is a full-time professor of Obuda University, Budapest, Hungary. He is teaching algorithms, metaheuristics, and GPU programming. His research areas are image processing, parallel programming, and metaheuristics. He is the head of the Hungarian High-Performance Computing Society.