Mathematical models of real-world problems in uncertainty quantification, optimisation and sensitivity analysis use a large number of parameters which can be expensive and time consuming. Often simulations have to be run multiple times to effectively study inputs and outputs. However, each simulation can be extremely expensive. An active subspace approach allows us to identify an important linear combination of parameters instead of analysing all of them individually. This can significantly reduce the dimension of the parameter study. We develop these ideas in a network science context to find important edges in a given network. Then, we make a comparison with the Sobol method, which is an alternative approach to sensitivity analysis. Moreover, we apply the active subspace method to examples in measurement science.Our analysis shows that the active subspace method on networks was able to identify important edge(s) or connections in a given graph. The active subspace method and Sobol method result in similar findings; however, the active subspace method is computationally less expensive. We are able to validate some of these results using extra information about the networks. In the measurement science setting, the results of the active subspace method match with the results obtained previously by NPL.
| Date of Award | 13 May 2022 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | University of Strathclyde |
|---|
| Supervisor | Alison Ramage (Supervisor) & Desmond Higham (Supervisor) |
|---|