Reliability-aware evolutionary neural architecture search under noisy fitness evaluation for generative adversarial networks

Weight-sharing evolutionary neural architecture search for generative adversarial networks (GANs) suffers from unreliable inherited-weight fitness, leading to search—retraining inconsistency and unstable evolutionary selection. This paper proposes Reliability-aware Memory-guided Evolutionary GAN Neural Architecture Search (RME-GAN), a reliability-driven evolutionary architecture search framework for noisy inherited-weight evaluation. Specifically, RME-GAN formulates evolutionary architecture search as a reliability-guided candidate refinement process, where imperfect inherited-weight fitness is evaluated with an explicit reliability signal and exploited through historical memory and feedback sampling. Within this framework, reliability-aware candidate evaluation and intra-front selection suppress the influence of short-term fitness fluctuations on evolutionary selection; persistent Subnetwork Memory serves as a replay mechanism for reliable historical candidates to preserve stable evolutionary search trajectories; and annealed three-route sampling with budget-annealed evolutionary search further balances exploration, exploitation, and model compactness. Experiments on CIFAR-10, STL-10, CIFAR-100, and ImageNet32 show that RME-GAN discovers compact and transferable generator architectures. It achieves an FID of 8.17 and an IS of 8.91 ± 0.08 on CIFAR-10, and obtains an FID of 18.67 and an IS of 10.27 ± 0.12 after transfer to STL-10. These results demonstrate that RME-GAN improves search stability, generation quality, and cross-dataset transferability, validating the effectiveness of reliability-aware evolutionary search under noisy inherited-weight evaluation.