We are in the process of constructing a Machine Learning (ML) agent for PDF determination.
The motivation is to arrive at a PDF determination which is optimized without any bias in the optimization choices. This goes beyond the previous NNPDF methodology in three respects. First, a single AI agent will replace a Monte Carlo set of neural networks trained to data. Second, the optimization of this agent will not rely on tuning a specific neural architecture and a genetic algorithm used for its minimization. Third, the validation of the method will not only be done a posteriori through closure tests but also will be an outcome of the learning process.
Several significant results have been achieved already. We have redesigned our main PDF determination tool as a modular structure based on a machine learning paradigm, using Keras and TensorFlow. Genetic minimization has been replaced with gradient descent. Hyperoptimization of the methodology based on K-folding is now possible and is being used to produce the next generation PDF sets.
Current directions of research involve the exploration of deeper figures of merit, the development of generalization tools in order to reliably extrapolate where there is no data, and the exploration of reinforcement learning as a way to add a further level of meta-optimization.