Automating GAN Hyperparameter Selection: Insights from The PSO and ABC Techniques

Generative Adversarial Networks (GANs) are powerful deep learning models that generate high-quality synthetic data for picture synthesis, ‎data augmentation, and video production. However, training GANs is difficult and resource-intensive because of non-convergence, mode ‎collapse, instability, and hyperparameter sensitivity. This research optimizes hyperparameters utilizing optimization methods to improve ‎GAN, with an experiment applying to digit dataset generation. Hybrid Particle Swarm Optimization (PSO) and Artificial Bee Colony (ABC) ‎algorithms were presented to automate hyperparameter tweaking, focusing on Training Epochs, Batch Size, and Label Smoothing. This ‎work introduced binarizing feature candidate selection instead of discrete values for optimization. This method represents parameters in ‎binary form, where 1 signifies selection and 0 indicates non-selection, for more efficient parameter space search. Through experimental ‎findings, the proposed technique significantly improves GAN training stability and output quality. Training for 25 epochs, with a batch size ‎of 128 and a label smoothing of 0.8, yields the most consistent and high-quality outcomes, with ABC surpassing PSO. This study ‎underscores the effect of hyperparameter optimization for enhancing GAN performance and introduces an innovative way for adjusting ‎critical parameters. Optimization methods like PSO and ABC can improve GAN training and support their use in digit generation and other ‎applications‎.