Research on Deep Neural Network Hyperparameter Optimization Method Based on Improved Tree Seed Algorithm

The performance of deep neural networks greatly depends on how their hyperparameters are set. However, tuning these hyperparameters by hand takes much time and is not efficient. To solve this problem, this paper presents a method for deep neural network hyperparameter optimization. This method uses an improved Tree Seed Algorithm (GTSA). Standard TSA has drawbacks in the complex and high-dimensional hyperparameter search space. We introduced a self-adaptive nonlinear step size dynamic adjustment mechanism and a dynamic seed number method based on multiple distributions. The first mechanism aims to balance the algorithm's global search and local search abilities. The second mechanism aims to increase population diversity and help the algorithm escape local optima. To test GTSA's effectiveness, we first evaluated its performance on the CEC 2014 test suite. The experimental results showed that GTSA has superior optimization performance on various benchmark functions. Then, we applied GTSA to optimize the hyperparameters of a convolutional neural network on the CIFAR-10 dataset. Compared to methods like Random Search, Bayesian Optimization, and Particle Swarm Optimization, GTSA significantly improved test accuracy and optimization efficiency. The experimental results prove that GTSA is an effective and competitive method for deep neural network hyperparameter optimization.