Submodular optimization and interpretable machine learning
Submodular functions are used to characterize the diminishing-returns property, which appears in many application areas, including information summarization, sensor placement, viral marketing, and more. Optimizing submodular functions has a rich history in mathematics and operations research, while recently, the subject has received increased attention due to the prevalent role of submodular functions in a broad range of data-science applications. In this talk we will discuss two recent projects on the topic of interpretable classification, both of which make interesting connections with submodular optimization. For the first project, we address the problem of multi-label classification via concise and discriminative rule sets. Submodularity is used to account for diversity, which helps avoiding redundancy, and thus, controlling the number of rules in the solution set. In the second project we aim to find accurate decision trees that have small size, and thus, are interpretable. We study a general family of impurity functions, including the popular functions of entropy and Gini-index, and show that a simple enhancement, relying on the framework of adaptive submodular ranking, can be used to obtain a logarithmic approximation guarantee on the tree complexity.
|
Portwings: Decoding the physics of flapping flight
In this seminar an overview will be given of the ERC AdG PortWings project and its achievements. The overview will first describe the initial goal and the context in which the project came to existence, and then will give attention to the various aspects of the results which have been achieved in the last years. This will include completely new methods to model and interconnect any multi-physical systems including the coupling of Navier-Stokes descriptions of fuids and non linear elasticity. Thanks to the new methodologies, novel numerical discretisation methods have been found which allow structure preserving integration of physical descriptions. This opens new possibilities on model reduction and control schemes able to use physically based models.
|
Optimal Predictive Control of Smart Cyber-physical Systems
The talk will discuss about the application of optimal control and model predictive control (MPC) algorithms in key problems from the power systems, smart cities, and assembly line domains. Perhaps not surprisingly, although arising from different application domains, the tackled problems have very similar mathematical descriptions, since the involved systems can be modelled as systems integrating dynamics, logics, and constraints, leading to models which can be easily integrated into discrete-time MPC formulations (often based on mixed-integer programming) or continuous time optimal control problems. The talk will present basic problem formulations and results for smart charging, tasks scheduling and traffic control problems, and will further outline the ongoing research efforts to make the control problems better scale to large scenarios. Challenges concerning arising cyber-physical security issues in the smart grid will be also briefly discussed.
|
Transformative Robotic Methodologies in Healthcare
The integration of robotics into healthcare has ushered in a new era of precision, efficiency, and improved patient outcomes. This talk explores the transformative role of control theory and advanced robotic methodologies in modern healthcare, including interaction force estimation and control, dynamic simulation, haptics and haptic rendering, and state estimation of infinite-dimensional systems. After providing a general overview of the topic and highlighting my contributions to the field, the presentation will offer a detailed analysis of the design and implementation of an innovative robotic system for superficial hyperthermia, from concept to experimental validation. Additionally, ongoing research related to this system will be discussed, focusing on the development of non-invasive temperature estimation of human body internal tissues through the combination of control theory and artificial intelligence methods.
|
Analysis and Synthesis of Emergent Behavior of Multi-agent Systems via Blended Dynamics Theorem
When a diverse set of dynamical systems are compelled to synchronize under a coupling condition, they display emergent behavior characterized by the blended dynamics. This phenomenon is crucial in designing heterogeneous multi-agent systems, where each agent collaborates towards a goal in a coordination. Such multi-agent systems or algorithms benefit from stability exchange among agents, an initialization-free nature allowing seamless integration or disengagement of agents, and resilience to production flaws and disturbances in node dynamics. Additionally, for each agent to undertake distinct tasks while aligning with others for a collective objective, certain internal variables of each agent must reach consensus, and the agreed-upon value should reflect some information about individual agents. We substantiate these points by demonstrating a few concrete examples.
|