Hyperparameter Optimization of Convolutional Neural Networks (CNN) using an Adaptive Particle Swarm Optimization (PSO) and Genetic Algorithm (GA)

Efficient hyperparameter optimization is critical for enhancing the performance of Convolutional Neural Networks (CNNs). However, existing approaches—including Particle Swarm Optimization (PSO), Genetic Algorithms (GA), and their hybrids—often suffer from premature convergence, poor adaptability, and inadequate exploration in high-dimensional search spaces, limiting their effectiveness in real-world applications such as real-time image processing, robotics, and autonomous systems. This research proposes PSO-CMLF-GA, a novel adaptive metaheuristic framework that integrates Adaptive Particle Swarm Optimization (PSO), Cauchy-Mutated Lévy Flight (CMLF), and a feedback-driven Genetic Algorithm (GA) under a unified, dual-trigger control architecture. Unlike conventional static hybrids, the proposed framework integrates real-time feedback, linearly decreasing strategies, and condition-triggered mechanisms to dynamically balance exploration and exploitation. The CMLF component enhances global and local search through coordinated adaptivity, while the GA phase is adaptively configured based on diagnostic feedback from the PSO phase, enabling a context-sensitive transition from global to local search. The optimization process operates in two phases. In the PSO phase, dynamically adjusted parameters guide the Particle Swarm Optimization (PSO) process and the baseline Lévy–Cauchy mechanisms. PSO’s adaptive response is triggered under mild conditions, while the adaptive mechanisms of Lévy flight with Cauchy mutation, are activated under more severe conditions. In the GA phase, crossover and mutation rates are dynamically adjusted using the same condition-driven framework. This cross-phase coordination and co-regulated search behaviour address the instability, inefficiency, and rigidity of traditional PSO-GA models. The proposed algorithm was validated using benchmark test functions and real-world bearing fault data, and its performance was compared against seven hybrid PSO-GA variants and seven metaheuristic algorithms, including advanced metaheuristics. Experimental results show that the proposed approach significantly enhances convergence speed, robustness, stability, and overall optimization performance. By addressing the limitations of traditional PSO, GA, and hybrid PSO-GA methods, this research offers a scalable, adaptive solution for efficient CNN hyperparameter tuning.