VASA: A Novel Adaptive Simulated Annealing Strategy for CNN Hyperparameter Optimization
Arab, Hamza, Mahraz, Abdelrahman, Boukouffallah, Abdallah, Bouyakoub, Rayane, Djeghri, Lotfi, Alismail, Dyna Hayem · Zenodo (CERN European Organization for Nuclear Research) · 2025
In this study, VASA is proposed as a hyperparameter optimization approachgrounded in an adaptive simulated annealing framework. The method applies a sepa-rate adaptive search process to each hyperparameter, guided by a weighted mechanismthat reflects the relative influence of each parameter. To help balance exploration andexploitation, VASA incorporates elite solutions and a tabu list, aiming to steer the searcheffectively while avoiding redundant configurations. Rather than depending on fixedannealing schedules or static heuristics, the approach offers a flexible framework intendedto support more efficient search dynamics in high-dimensional spaces.