Hybrid adaptive Wolf-Particle swarm optimization algorithm and its application in CNN neural network hyperparameters optimization

This paper introduces the Hybrid Adaptive Wolf-Particle Swarm Optimization (HAWPSO) algorithm, a novel metaheuristic framework designed to address the intricate challenge of optimizing hyperparameters in convolutional neural networks (CNNs) within dynamic and high-dimensional search spaces. Conventional gradient-based optimization methods frequently suffer from premature convergence and suboptimal performance, especially in non-convex loss landscapes. To overcome these constraints, HAWPSO synergistically merges the global exploratory power of Particle Swarm Optimization (PSO) with the structured, hierarchical local search capabilities of the Grey Wolf Optimizer (GWO). The algorithm integrates four key innovative elements: a dual-population co-evolution mechanism, a nonlinear adaptive inertia weight strategy, a quantum tunneling operator for escaping local optima, and a rigorous convergence proof grounded in Markov chain theory. Extensive experimental validation on 38 benchmark functions, including those from the CEC2017 special session, confirms that HAWPSO surpasses established algorithms such as PSO, GWO, Whale Optimization Algorithm (WOA), and Sparrow Search Algorithm (SSA) in terms of convergence speed and solution accuracy. Notably, HAWPSO performs exceptionally well in managing multimodal and fixed-dimension multimodal functions. When applied to the CIFAR-10 dataset, the algorithm optimizes five critical CNN hyperparameters, enhancing classification accuracy from 59.6% to 77.6%. This study therefore presents a robust, adaptive, and theoretically grounded optimization framework, well-suited for complex, non-convex problems in machine learning and beyond.