Learning for Level Set estimation in Simulation-Based Inference
Simulators are ubiquitous in applied sciences. We propose “Excursion” to accelerate an analysis involving a simulator while optimizing costs. Excursion is a tool to smartly estimate level sets of computationally expensive black box functions. It is a confluence of Gaussian Process Regression and Active Learning. The tool queries a black box to perform experiments (choice of parameters) in order to maximize a quantity close to the total information gain. Excursion uses GPyTorch with state-of-the-art fast posterior fitting techniques and takes advantage of GPUs to scale computations to higher dimensions.
In this talk, we demonstrate that Excursion significantly outperforms traditional grid search approaches and we explain the ongoing efforts on improving pMSSM scans at ATLAS.