A Review on Analyzing Grokking Dynamics in Swarm-Based Metaheuristic Optimization for Behavioural Disorder Classification.

The classification of behavioural disorders is difficult, due to smaller dataset of patients, a large feature space dimension, and high clinical variability, which may lead to low model generalization. Although feature selection and hyper parameter tuning are the most popular applications of Swarm Intelligence (SI) algorithms, the majority of current methods stop the optimization process as soon as the optimum has been reached. On the other hand, recent research into gradient-based learning has demonstrated the appearance of a phenomenon they refer to as grokking, where long optimization training can cause late generalization. This review examines whether similar post-convergence behaviour could be relevant to non-gradient, population-based optimization. A literature review was carried out through a structured search of scientific databases from 2004 to 2026. The literature thus selected was examined in terms of optimization strategy, learning behaviour, characteristics of the dataset, and clinical applicability. The review offers four main research gaps: (1) post-convergence analysis missing from swarm intelligence algorithms, (2) insufficient knowledge of grokking in decentralized optimization, (3) lack of research in the field of behavioural healthcare, and (4) no validation performed on noisy clinical datasets. From these observations, a conceptual framework for using grokking as a link between swarm-based optimization is suggested. The review serves as a basis for further research on post-convergence optimization, and could be useful for the development of more stable and reliable feature selection methods for clinical decision supporting systems.