Optical neural networks (ONNs) offer high-speed, energy-efficient computation but face training challenges due to slow convergence and local optima. To address this, we propose the Elite Preserving and Local Tuning Genetic Algorithm (ELGA), tailored for ONNs. ELGA incorporates a residual-based population competition strategy and a firefly-inspired local refinement mechanism. Integrated into a Mach–Zehnder interferometer (MZI)-based ONNs architecture, ELGA is evaluated on five benchmark datasets. Results show improved classification accuracy and faster convergence over stand-ard GA, FA, and M2SGA, demonstrating enhanced stability and suitability for real-time ONNs training in re-source-constrained scenarios.