HBO-NAS: class-aware zero-cost fitness for diversity-preserving neural architecture search through hybrid breeding optimization algorithm

Abstract

Neural Architecture Search (NAS) with evolutionary computing increasingly relies on zero-cost proxies to mitigate the prohibitive computational cost of training candidate networks. However, existing proxies are mainly designed to maximize score–accuracy correlation, neglecting the landscape structure of objectives required to sustain population diversity. To address this issue, a class-aware, training-free objective function is proposed, which utilizes intra-class compactness and inter-class separability to induce a structured framework, a multi-modal fitness landscape that naturally prevents premature convergence. This capability effectively facilitates the discovery of a broader range of different, high-performing structures. When evaluated using Hybrid Breeding Optimization algorithm, our method consistently yields superior optimization performance, achieving the average accuracy of 71.18% on ImageNet16-120 within the DARTS search space, which is nearly equal to the reported best-performing architecture with the accuracy of 72.00%, while maintaining a high level of population diversity. These findings show the critical shift towards a search-centric perspective, where shaping the landscape structure of objectives is as important as ranking fidelity for discovering diverse structures.