HiveNAS: Neural Architecture Search Using Artificial Bee Colony Optimization

Traditionally, constructing Neural Networks has depended on specialized expertise and on iterative, intuition-driven optimization. Neural Architecture Search (NAS) frameworks were introduced to systematically scan the architectural serach space and automate Neural Network construction. Although some optimization schemes, such as Genetic Algorithms, have been studied in depth within NAS, many Metaheuristic Optimization methods have received comparatively less attention. In this study, we investigate the suitability of Artificial Bee Colony optimization as the search engine for NAS. The proposed framework, HiveNAS, attains stronger results than representative state-of-the-art Swarm Intelligence-based NAS frameworks in a fraction of the time, while also encouraging efficient parallel exploration of the search space. We additionally formalize a discrete NAS adaptation of ABC through an explicit neighbor operator and a low-fidelity, compute-aware evaluation protocol, enabling favorable accuracy and efficiency trade-offs under tight computational budgets.