Hyperparameter Optimization of Generative Adversarial Network Models for High-Energy Physics Simulations

Event generation and detector simulation are essential for physics analyses at the Large Hadron Collider, but are also computationally expensive. Different Machine Learning-based generative models are exploited to reduce the computational cost. Those generative models can be classified into four categories: (1) Variational Autoencoders [1], which learn a stochastic map from the data space to a latent space and back, preserving the statistics of the latent space and data space; (2) Normalizing Flows [2] use invertible transformations so that the probability density can be computed and the generator is optimized using the log likelihood; (3) score-based generative models [3, 4], which generate samples from noise by repeatedly perturbing the data with a diffusion equation, and learning to reverse the perturbation via estimating the diffusion function; (4) Generative Adversarial Networks (GAN) [5], which optimize the generator network by means of an auxiliary network ('discriminator') that tries to classify generated examples from real examples. GAN will remain as an important generative model in High Energy Physics because of its unique features. GANs have been used in many aspects of High Energy Physics to accelerate computationally intensive physics simulations.