DOI: 10.54569/aair.1960245 ISSN: 2757-7422

A Multi-Stage Metaheuristic Framework for Reliable WGAN-GP Hyperparameter Optimization

Emre Mert Dübüş, Yusuf Sönmez, Hamdi Tolga Kahraman
Show PDF Cite
Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) improve adversarial training stability, but performance remains highly sensitive to hyperparameters governing generator–critic balance, learning rates, optimizer momentum, regularization, and update frequency. This study presents a progressive three-stage metaheuristic framework for WGAN-GP hyperparameter optimization. Stage 1 performs low-cost preliminary exploration on MNIST. Stage 2 refines the FFHQ 64 × 64 search space using a baseline and LSHADE-cnEpSin. Stage 3 compares Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Differential Evolution (DE), and LSHADE-cnEpSin under equal function-evaluation budgets with real Fréchet Inception Distance (FID) as the primary metric. The main result is a search-validation gap: PSO achieves the best short-horizon search-phase FID, whereas GA achieves the best long-horizon final-validation FID. GA is therefore interpreted as showing delayed exploitation in the reported setting, not universal superiority. LSHADE-cnEpSin obtains the lowest search-phase mean FID and stable candidate-level behavior, showing that population-level consistency and final best-candidate quality are distinct evaluation dimensions. These results support staged, fair-budget validation as a more reliable protocol for WGAN-GP hyperparameter optimization.

More from our Archive