Abstract Integrating deep learning methods into metaheuristic algorithms has gained attention for addressing design-related issues and enhancing performance. The primary objective is to improve solution quality and convergence speed within solution search spaces. This study investigates the use of deep learning methods as a generative model to learn historical content, including global best and worst solutions, solution sequences, function evaluation patterns, solution space characteristics, population modification trajectories, and movement between local and global search processes. An LSTM-based architecture is trained on dynamic optimization data collected during the metaheuristic optimization process. The trained model generates an initial solution space and is integrated into the optimization algorithms to intelligently monitor the search process during exploration and exploitation phases. The proposed deep learning-based methods are evaluated on 55 benchmark functions of varying complexities, including CEC 2017 and compared with 13 biology-based, evolution-based, and swarm-based metaheuristic algorithms. Experimental results demonstrate that all the deep learning-based optimization algorithms achieve high-quality solutions, faster convergence rates, and significant performance improvements. These findings highlight the critical role of deep learning in addressing design issues, enhancing solution quality, trajectory, and performance speed in metaheuristic algorithms.

Abstract For last two decades, nature‐inspired metaheuristic algorithms together with their modified, improved, and hybrid versions have been gaining huge popularity in the field of optimization in solving continuous and complex real‐life optimization problems. In this work, a novel improved symbiosis organism search (SOS) algorithm, called self‐adaptive beneficial factor‐based improved SOS (SaISOS, in short) is suggested. The self‐adaptive benefit factors and a modified mutualism phase (called “Three‐way mutualism phase”) have been introduced here to upgrade the performance of SOS algorithm. A random weighted reflection coefficient and a new control operator have also been introduced. To validate the proposed algorithm and to compare its performance with other state‐of‐the‐art algorithms, 15 IEEE‐CEC 2015 functions have been employed and the experimental results confirm that SaISOS provides competitive results on most occasions. Also, the proposed algorithm is used to solve five real‐world optimization problems. Considering the average output, it is observed that the proposed method performs significantly better in solving the real‐world problems compared to the alternative state‐of‐the art techniques considered in this work.

The search for food stimulated by hunger is a common phenomenon in the animal world. Mimicking the concept, recently, an optimization algorithm Hunger Games Search (HGS) has been proposed for global optimization. On the other side, the Whale Optimization Algorithm (WOA) is a commonly utilized nature-inspired algorithm portrayed by a straightforward construction with easy parameters imitating the hunting behavior of humpback whales. However, due to minimum exploration of the search space, WOA has a high chance of trapping into local solutions, and more exploitation leads it towards premature convergence. The concept of hunger from HGS is merged with the food searching techniques of the whale to lessen the inherent drawbacks of WOA. Two weights of HGS are adaptively designed for every whale using the respective hunger level for balancing search strategies. Performance verification of the proposed hunger search-based whale optimization algorithm (HSWOA) is done by comparing it with 10 state-of-the-art algorithms, including three very recently developed algorithms on 30 classical benchmark functions. Comparison with some basic algorithms, recently modified algorithms, and WOA variants is performed using IEEE CEC 2019 function set. Statistical performance of the proposed algorithm is verified with Friedman's test, boxplot analysis, and Nemenyi multiple comparison test. The operating speed of the algorithm is determined and tested with complexity analysis and convergence analysis. Finally, seven real-world engineering problems are solved and compared with a list of metaheuristic algorithms. Numerical and statistical performance comparison with state-of-the-art algorithms confirms the efficacy of the newly designed algorithm.