Machine learning aided metaheuristics: A comprehensive review of hybrid local search methods

Abstract

Machine learning-based methods have emerged as competitors to traditional metaheuristic-based solutions in many areas. Besides investigating their effectiveness, it raises the question of whether these methods can be combined. This paper presents a systematic literature review based on P.R.I.S.M.A. methodology to provide a state-of-the-art overview of machine learning-assisted metaheuristics, focusing on local search algorithms such as Hill Climbing, Tabu Search, and Simulated Annealing. The review is based on a comprehensive evaluation of 48 related articles. These studies illustrate the most common applications of hybrid methods in various fields, including physical simulations and scheduling problems. This paper demonstrates commonly used assembly options, such as metamodeling and machine learning aided initialization, along with some novel ideas like early stopping and cooling control based on neural networks. The evaluation of the results reveals several potential machine learning methods, such as Deep Neural Networks, Hopfield Networks, and Self-Organizing Maps, to assist the metaheuristics. Different training methods for these approaches, including online vs offline training and sources of training data, are also reviewed. Most papers address real-world problems, but there are several intriguing ideas for improving local searches in general. • A P.R.I.S.M.A. review of machine learning assisted metaheuristic methods is presented. • The study focuses on local search hybrids such as Tabu Search and Simulated Annealing. • Several potential machine learning methods are revealed (DNN, Hopfield, SOM, etc.). • Commonly used assembly options, such as metamodeling, are discussed. • Different available training methods, both online and offline, are reviewed.

Topics & Concepts