|
|
|
|
|
|
| Sep 26 |
Thu |
UGRI-SURE students |
SIAM-IMA Chapter Seminar |
| 13:30 |
|
|
|
|
LT5, Hicks |
|
|
|
|
| Sep 27 |
Fri |
UGRI-SURE students |
SIAM-IMA Chapter Seminar |
| 13:30 |
|
|
|
|
LT5, Hicks |
|
|
|
|
| Nov 6 |
Wed |
Dr Kieran Kalair (Smith Institute) |
SIAM-IMA Chapter Seminar |
| 14:00 |
|
Mathematics and explainable machine learning to support a zero-carbon grid.
|
|
|
K14 |
| |
Abstract:
The Smith Institute are a mathematics, data science and AI consultancy based in Oxford that solve practical, novel problems using these approaches. I will talk about what we do, and focus on a specific project we have been working on over the last 3 years – called Dynamic Reserve Setting – which looks at how much ‘spare capacity’ needs to be held to secure the GB energy grid. In this problem we care about explainability in our modelling, understanding extreme events in our data, and capturing non-linear relationships that exist in the energy grid. I will talk about both our modelling approach and how we feed results into practical usage. I will end with examples of other problems and methodologies we have worked on.
|
|
|
|
|
| Nov 20 |
Wed |
Dr Ross Drummond (University of Sheffield (EEE)) |
SIAM-IMA Chapter Seminar |
| 14:00 |
|
Robustifying neural networks: The view from Control Theory
|
|
|
K14, Hicks / Google Meet |
| |
Abstract:
This talk will describe some results on applying ideas from robust control theory to design and interpret neural networks. The basic idea is to use results from absolute stability theory to train neural networks with robustness guarantees built in. Furthermore, a method to represent model predictive control directly in terms of a neural network is also discussed. The overriding goal is to turn robustness analysis problems for neural networks into synthesis problems to overcome some of their limitations.
|
|
|
|
|
| Nov 27 |
Wed |
Pablo Rodolfo Baldivieso Monasterios (University of Sheffield (EEE)) |
SIAM-IMA Chapter Seminar |
| 14:00 |
|
A robust solution to the plug and play problem: a study of interconnectedness.
|
|
|
K14, Hicks / Google Meet |
| |
Abstract:
Networked systems are a collection of dynamical systems interconnected via physical or digital links. These networks are ubiquitous in the modern world and range from power networks to the financial system. A particular feature of these networks is their size changes over time: new elements are added, and some old elements are removed. The control of these systems presents a significant challenge. Classical techniques are monolithic, and changes in size or the interconnection structure may require a complete redesign of controllers. The plug-and-play paradigm necessitates controllers with reconfiguration capabilities to adapt to such structural changes.
In contrast, coalitional control offers a control methodology inherently capable of reconfiguration. In this talk, we discuss how the reconfiguration properties of a coalitional controller arise via partitioning the network's set of elements. We endow each of these elements with robust distributed predictive controllers designed to withstand interconnections through invariant sets. The size of the interconnections parameterises these invariant sets; the shrinkage of these sets allows each agent to tolerate exogenous disturbances. The addition of new elements to a network is modelled as a disturbance. Upon adding a new element, its neighbouring elements coalesce to manage transients. We show that this controller allows us to conclude the asymptotic stability of the equilibrium point and constraint satisfaction of the overall network.
|
|
|
|
|
| Dec 11 |
Wed |
Dr Carolanne Vouriot (University of Sheffield (School of Mechanical, Aerospace and Civil Engineering)) |
SIAM-IMA Chapter Seminar |
| 14:00 |
|
|
|
|
K14, Hicks |
|
|
|
|
