UC Irvine

Generative models with iterative methodologies are used to probe high-dimensional sub- spaces of the Minimal Super Symmetric Model (MSSM) in a variety of contexts. First, a dimensional reduction framework is used to sample theory points from a low-dimensional latent space and then de-code them back into theory parameters. Iterative density estimation sampling is used here to maximize efficiency as well as to filter by additional constraints post facto without further training. Next, an iterative usage of generative flows is used to produce models which can sample theory points under multiple simultaneous, without loss of search resolution. An initial flow is trained on a single constraint, after which a new dataset generated from that flow is used to create training data for the learning of two simultaneous constraints. This process is repeated up to seven simultaneous constraints, with high efficiency. Finally, we explore a new methodology which approximates an inverse mapping from experimental observable space to theory parameter space by learning to make small corrections to a ’bad guess’ candidate theory. Iterative corrections, taking the previous output theory point as a new input, are used to achieve sub-constraint-window accuracy. A secondary network is also trained to monitor the progress of these iterations in order to maximize efficacy.