The success of Convolutional Neural Networks (CNNs) in Facial Expression Recognition (FER) depends strongly on choosing the right hyperparameters. Manually tuning them is time-consuming and often ineffective, which is why metaheuristic optimizers have become an attractive alternative. In this study, we present a comparative analysis of ten well-known optimization methods—Genetic Algorithm (GA), Grey Wolf Optimizer (GWO), Squirrel Search Optimization (SSO), Quantum Gazelle Optimization (QGOA), Red Fox Optimization (RFO), Adaptive Search Optimizer (ASO), Football Team Training Algorithm (FTTA), exhaustive Grid Search, and a baseline CNN. Each approach was tested on the CK+, JAFFE and FER2013 dataset under ideal conditions. Among them, TPE, ASO achieved the highest accuracy, reaching 98.92%. To further evaluate robustness, we also tested the models on a version of the dataset augmented with synthetic noise. The results showed clear differences in performance: some optimizers excelled on clean data but struggled once noise was introduced, while others generalized better under imperfect conditions. These findings underline the importance of considering not only peak accuracy but also noise resilience when selecting optimizers for real-world FER systems.