Improvement strategies for heuristic algorithms based on machine learning and information concepts: a review of the seahorse optimization algorithm

To overcome the mechanical limitations of traditional inertia weight optimization methods, this study draws inspiration from machine learning models and proposes an inertia weight optimization strategy based on the K-nearest neighbors (KNN) principle with dynamic adjustment properties. Unlike conventional approaches that determine inertia weight solely based on the number of iterations, the proposed strategy allows inertia weight to more accurately reflect the relative distance between individuals and the target value. Consequently, it transforms the discrete “iteration-weight” mapping ($t\rightarrow w$) into a continuous “distance-weight” mapping ($d\rightarrow w$), thereby enhancing the adaptability and optimization capability of the algorithm. Furthermore, inspired by the entropy weight method, this study introduces an entropy-based weight allocation mechanism in the crossover and mutation process to improve the efficiency of high-quality information inheritance. To validate its effectiveness, the proposed strategy is incorporated into the Seahorse Optimization Algorithm (SHO) and systematically evaluated using 31 benchmark functions from CEC2005 and CEC2021 test suites. Experimental results demonstrate that the improved SHO algorithm, integrating the logistic-KNN inertia weight optimization strategy and the entropy-based crossover-mutation mechanism, exhibits significant advantages in terms of convergence speed, solution accuracy, and algorithm stability. To further investigate the performance of the proposed improvements, this study conducts ablation experiments to analyze each modification separately. The results confirm that each individual strategy significantly enhances the overall performance of the SHO algorithm.