This project focuses on the Uncertainty Quantification (UQ) in network throughput predictions. Scientific user facilities generate massive datasets, and remote data transfers play a key role in fields like climate modeling, bioinformatics, and particle physics. Effective UQ can help allocate network and storage resources more efficiently.
The project aims to develop and evaluate UQ methods, including Bootstrap, Conformal, and Predictability, Computability, and Stability (PCS) framework, to provide accurate perturbation intervals for evaluating future data transfers.
- UQ Techniques: Bootstrap, Conformal, PCS Framework.
- Machine Learning: Linear, Bagging, Boosting, and Neural Nets.
- Evaluation: Coverage, interval width, prediction accuracy.
The dataset will be sourced from Tstat logs at NERSC (National Energy Research Scientific Computing Center), providing detailed information on network transfers