Posted September 14, 2022
Last modified September 27, 2022
Control and Optimization Seminar Questions or comments?
9:30 am – 10:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Maryam Yashtini, Georgetown University
Counting Objects by Diffused Index: Geometry-Free and Training-Free Approach
Counting objects is a fundamental but challenging problem. In this talk, we propose diffusion-based, geometry-free, and learning-free methodologies to count the number of objects in images. The main idea is to represent each object by a unique index value regardless of its intensity or size, and to simply count the number of index values. First, we place different vectors, referred to as seed vectors, uniformly throughout the mask image. The mask image has boundary information of the objects to be counted. Secondly, the seeds are diffused using an edge-weighted harmonic variational optimization model within each object. We propose an efficient algorithm based on an operator splitting approach and alternating direction minimization method, and theoretical analysis of this algorithm is given. An optimal solution of the model is obtained when the distributed seeds are completely diffused such that there is a unique intensity within each object, which we refer to as an index. For computational efficiency, we stop the diffusion process before a full convergence, and propose to cluster these diffused index values. We refer to this approach as Counting Objects by Diffused Index (CODI). We explore scalar and multi-dimensional seed vectors. For scalar seeds, we use Gaussian fitting in a histogram to count, while for vector seeds, we exploit a high-dimensional clustering method for the final step of counting via clustering. The proposed method is flexible even if the boundary of the object is not clear nor fully enclosed. We present counting results in various applications such as biological cells, agriculture, concert crowds, and transportation. Some comparisons with existing methods are presented.
Posted December 12, 2022
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Giulia Giordano, University of Trento, Italy
SIAG on Control and Systems Theory Prize Awardee
What We Can Learn from the System Structure in Biology and Epidemiology
Biological, ecological and epidemiological systems can be seen as dynamical networks, namely dynamical systems that are naturally endowed with an underlying network structure, because they are composed of subsystems that interact according to an interconnection. Despite their large scale and complexity, natural systems often exhibit extraordinary robustness that preserves fundamental properties and qualitative behaviors even in the presence of huge parameter variations and environmental fluctuations. First, we focus on biochemical reaction networks and look for the source of the amazing robustness that often characterizes them, by identifying properties and emerging behaviors that exclusively depend on the system structure (i.e., the graph structure along with qualitative information), regardless of parameter values. We introduce the BDC-decomposition to capture the system structure and enable the parameter-free assessment of important properties, including the stability of equilibria and the sign of steady-state input-output influences, thus allowing structural model falsification and structural comparison of alternative mechanisms proposed to explain the same phenomenon. Then, inspired by the COVID-19 pandemic and the observation that compartmental models for epidemics can be seen as a special class of chemical reaction networks, we consider epidemiological systems describing the spread of infectious diseases within a population, along with control approaches to curb the contagion. We illustrate strategies to cope with the deep uncertainty affecting parameter values and optimally control the epidemic.
Posted January 27, 2023
Last modified January 29, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Raphael Jungers, Université Catholique de Louvain
Data-Driven Control of Hybrid Systems and Chance-Constrained Optimization
Control systems are increasingly complex, often to the point that obtaining a model for them is out of reach. In some situations, (parts of) the systems are proprietary, so that the very equations that rule their behavior cannot be known. On the other hand, the ever-growing progress in hardware technologies often enables one to retrieve massive data, e.g., from embedded sensors. Due to these evolutions, control theory can alternatively be viewed as a model-free and data-driven paradigm. For linear time-invariant systems, classical results from identification theory provide a rather straightforward approach. However, these approaches become least inefficient if one relaxes the assumptions they rely upon, e.g., linearity, Gaussian noise, etc. This is especially the case in safety-critical applications, where one needs guarantees on the performance of the obtained solution. Despite these difficulties, one may sometimes recover firm guarantees on the behavior of the system. This may require changing one's point of view on the nature of the guarantees we seek. I will provide examples of such results for different control tasks and different complex systems, and will raise the question of theoretical fundamental barriers for these problems.
Posted December 12, 2022
Last modified February 8, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Patrick L. Combettes, North Carolina State University
IEEE Fellow
Perspective Functions
I will discuss mathematical and computational issues pertaining to perspective functions, a powerful concept that makes it possible to extend a convex function to a jointly convex one in terms of an additional scale variable. Recent results on perspective functions with nonlinear scales will also be discussed. Applications to inverse problems and statistics will also be presented.
Posted February 3, 2023
Last modified February 6, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Enrique Zuazua, Friedrich-Alexander-Universität Erlangen-Nürnberg
2022 SIAM W.T. and Idalia Reid Prize Winner
Control and Machine Learning
In this lecture, we present some recent results on the interplay between control and machine learning, and more precisely, supervised learning and universal approximation. We adopt the perspective of the simultaneous or ensemble control of systems of residual neural networks (or ResNets). Roughly, each item to be classified corresponds to a different initial datum for the Cauchy problem of the ResNets, leading to an ensemble of solutions to be driven to the corresponding targets, associated to the labels, by means of the same control. We present a genuinely nonlinear and constructive method, allowing us to show that such an ambitious goal can be achieved, estimating the complexity of the control strategies. This property is rarely fulfilled by the classical dynamical systems in mechanics and the very nonlinear nature of the activation function governing the ResNet dynamics plays a determinant role. It allows deformation of half of the phase space while the other half remains invariant, a property that classical models in mechanics do not fulfill. The turnpike property is also analyzed in this context, showing that a suitable choice of the cost functional used to train the ResNet leads to more stable and robust dynamics. This lecture is inspired in joint work, among others, with Borjan Geshkovski (MIT), Carlos Esteve (Cambridge), Domènec Ruiz-Balet (IC, London) and Dario Pighin (Sherpa.ai).
Posted January 17, 2023
Last modified February 27, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Spring Berman, Arizona State University
DARPA Young Faculty and ONR Young Investigator Awardee
Scalable Control of Robotic Swarms with Limited Information
Robotic swarms are currently being developed for many applications, including environmental sensing, exploration and mapping, infrastructure inspection, disaster response, agriculture, and logistics. However, significant technical challenges remain before they can be robustly deployed in uncertain, dynamic environments. We are addressing the problem of controlling swarms of robots that lack prior data about the environment and reliable inter-robot communication. As in biological swarms, the highly resource-constrained robots would be restricted to information obtained through local sensing and signaling. We are developing scalable control strategies that enable swarms to operate largely autonomously, with user input consisting only of high-level directives that map to a small set of robot parameters. In this talk, I describe control strategies that we have designed for collective tasks that include coverage, mapping, and cooperative manipulation. We develop and analyze models of the swarm at different levels of abstraction based on differential equations, Markov chains, and graphs, and we design robot controllers using feedback control theory and optimization techniques. We validate our control strategies in simulation and on experimental test beds with small mobile robots.
Posted November 30, 2022
Last modified February 28, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Jacquelien Scherpen, University of Groningen
IEEE Fellow, Automatica Best Paper Prize Awardee
Model Reduction for Nonlinear Control Systems Based on Differential Balancing and Data
We present the standard balancing theory for nonlinear systems, which is based on an analysis around equilibrium points. Its extension to the contraction framework offers computational advantages, and is presented as well. We provide definitions for controllability and observability functions and their differential versions which can be used for simultaneous diagonalization procedures, providing a measure for importance of the states, as can be shown by the relation to the Hankel operator. In addition, we propose a data-based model reduction method based on differential balancing for nonlinear systems whose input vector fields are constants by utilizing its variational system. The difference between controllability and reachability for the variational system is exploited for computational reasons. For a fixed state trajectory, it is possible to compute the values of the differential Gramians by using impulse and initial state responses of the variational system. Therefore, differential balanced truncation is doable along state trajectories without solving nonlinear partial differential equations.
Posted February 14, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Frank Allgower, University of Stuttgart
IFAC Fellow
Data-driven Model Predictive Control: Concepts, Algorithms and Properties
While recent years have shown rapid progress of learning-based and data-driven methods to effectively utilize data for control tasks, providing rigorous theoretical guarantees for such methods is challenging and an active field of research. This talk will be about a recently developed framework for model predictive control (MPC) of unknown systems based only on input-output data which admits exactly such guarantees. The proposed approach relies on the Fundamental Lemma of Willems et al. which parametrizes trajectories of unknown linear systems using data. First, we cover MPC schemes for linear systems with a focus on theoretical guarantees for the closed loop, which can be derived even if the data are noisy. Building on these results, we then move towards the general, nonlinear case. Specifically, we present a data-driven MPC approach which updates the data used for prediction online at every time step and, thereby, stabilizes unknown nonlinear systems using only input-output data. In addition to introducing the framework and the theoretical results, we also report on successful applications of the proposed framework in simulation and real-world experiments.
Posted February 13, 2023
Last modified April 9, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Ningshi Yao, George Mason University
Resolving Contentions Through Real-Time Control and Scheduling for Cyber Physical Human Systems
Shared resources, such as cloud computing and communication networks, are widely used in large-scale real-time systems to increase modularity and flexibility. When multiple systems need to access a shared resource at the same time and the demands exceed the total supply, a contention occurs. A scheduling strategy is needed to determine which systems can access the resource first to resolve contentions. However, such a scheduling mechanism inevitably introduces time-varying delays and may degrade the system performance or even sabotage the stability of control systems. Considering the coupling between scheduling and control, this talk presents a novel sample-based method to co-design scheduling strategies and control laws for coupled control systems with shared resources, which aims to minimize the overall performance degradation caused by contentions. The co-design problem is formulated as a mixed integer optimization problem with a very large search space, rendering difficulty in computing the optimal solution. To solve this challenge, we describe a contention resolving model predictive control (CRMPC) method to dynamically design optimal scheduling and control in real-time. With fundamental assumptions in scheduling theory, the solution computed by CRMPC can be proved to be globally optimal. CRMPC is a theoretical framework that is general and can be applied to many applications in cyber-physical-human systems. The effectiveness of CRMPC has been verified in real-world applications, such as networked control systems, traffic intersection management systems, and human multi-robot collaboration systems. The performance of CRMPC was compared with well-known scheduling methods and demonstrated significant improvements.
Posted December 12, 2022
Last modified April 11, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Maria Elena Valcher, University of Padova
Fellow of IEEE and of IFAC
On the Influence of Homophily on the Friedkin-Johnsen Model
Over the last few decades, the modelling and analysis of sociological phenomena have attracted the interests of researchers from various fields, such as sociology, economics, and mathematics. Opinion dynamics models aim to describe and predict the evolution of the opinions of a group of individuals as a result of their mutual influence/appraisal. One of the most celebrated opinion dynamics models is the Friedkin-Johnsen (FJ) model, that captures the attitude of individuals to form their opinions by balancing exogenous and endogenous influences. On the one hand they value the opinions of the other individuals, weighted by the appraisals they have of them, and on the other hand they tend to adhere to their original opinions, that represent a permanent bias, to an extent that depends on the agent stubbornness. In the classical FJ model the weights that each agent gives to the opinions of the others are fixed. However, this is not consistent with other opinion dynamics models, where the weight matrix is time varying and it updates based on a homophily model: individuals decide which individuals they want to be influenced by (and on the contrary which individuals they want to distance their opinions from) based on the correlation between their opinion vectors. In this talk we will explore some recent results regarding this extended FJ model and present some future directions and challenges related to opinion dynamics models.
Posted January 17, 2023
Last modified April 16, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Weiwei Hu, University of Georgia
Optimal Control for Suppression of Singularity in Chemotaxis
In this talk, we discuss the problem of optimal control for chemotaxis governed by the parabolic-elliptic Patlak-Keller-Segel (PKS) system via flow advection. The main idea is to utilize flow advection for enhancing diffusion as to control the nonlinear behavior of the system. The objective is to determine an optimal strategy for adjusting flow strength for advection so that the local in time blow up of the solution can be suppressed. Rigorous proof of existence of an optimal solution and derivation of first-order optimality conditions for solving such a solution are presented. Numerical experiments based on 2D cellular flows in a rectangular domain are conducted to demonstrate our ideas and designs.
Posted January 26, 2023
Last modified April 2, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Wim Michiels, KU Leuven
Strong Relative Degree of Time-Delay Systems with Non-Commensurate Delays
The presentation addresses the notion of relative degree for linear time-delay systems of retarded type, when the common assumption of commensurate delays is dropped. Algebraic conditions are provided that fully exploit the delay dependence structure. It is shown that the relative degree may be sensitive to delay perturbations, which is the basis of a novel notion of relative degree, called strong relative degree. This notion is characterized algebraically and computationally in the SISO and MIMO settings. Using the obtained characterizations and a benchmark problem, which illustrates that invariant zeros may be characterized as zeros of quasi-polynomials of retarded, neutral or advanced type, light is shed on existence conditions of a normal form. The novel concepts and theoretical results also play a role in the design and analysis of extended PD controllers, as illustrated. Finally, connections are established with the notion of strong stability and strong H2-norm for delay equations of neutral type and delayed descriptor systems. This work is in collaboration with Bin Zhou from Harbin University of Technology.
Posted February 13, 2023
Last modified April 25, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Matthew Hale, University of Florida
AFOSR Young Investigator, ONR Young Investigator, and NSF CAREER Program Awardee
Resilient Multi-Agent Coordination: From Theory to Practice
A multi-agent system is any collection of decision-makers that collaborates on a common task. A distinguishing feature is that communications among agents provide the feedback signals needed for autonomous decision-making. For example, a team of drones may exchange location data and images to jointly map an area. There is now a large literature on multi-agent systems, though practical implementations are often fragile or only done in controlled environments. A fundamental challenge is that agents’ communications in realistic environments can be impaired, e.g., by delays and intermittency, and thus agents must rely on impaired feedback. To transition theory to practice, such systems need novel coordination techniques that are provably resilient to such impairments and validated in practice under realistic conditions. In this talk, I will cover two recent developments in my group that have successfully transitioned novel theory to practice for multi-agent systems facing asynchronous communications. The first considers a class of geometrically complex coordination tasks – namely those given by constrained nonconvex programs – and provides provable guarantees of performance that are borne out in practice onboard teams of drones. The second considers a class of time-varying task specifications for agents that can change unpredictably. Theoretical results show that agents can complete this class of task under mild restrictions, and validation is provided by a team of lighter-than-air agents in a contested environment.
Posted August 25, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Cristopher Hermosilla, Universidad Técnica Federico Santa María
Hamilton-Jacobi-Bellman Approach for Optimal Control Problems of Sweeping Processes
This talk is concerned with a state constrained optimal control problem governed by a Moreau's sweeping process with a controlled drift. The focus of this work is on the Bellman approach for an infinite horizon problem. In particular, we focus on the regularity of the value function and on the Hamilton-Jacobi-Bellman equation it satisfies. We discuss a uniqueness result and we make a comparison with standard state constrained optimal control problems to highlight a regularizing effect that the sweeping process induces on the value function. This is a joint work with Michele Palladino (University of L’Aquila, Italy) and Emilio Vilches (Universidad de O’Higgins, Chile).
Posted August 18, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Mario Sznaier, Northeastern University
IEEE Fellow, IEEE Control Systems Society Distinguished Member Awardee
Why Do We Need Control in Control Oriented Learning?
Despite recent advances in machine learning (ML), the goal of designing control systems capable of fully exploiting the potential of these methods remains elusive. Modern ML can leverage large amounts of data to learn powerful predictive models, but such models are not designed to operate in a closed-loop environment. Recent results on reinforcement learning offer a tantalizing view of the potential of a rapprochement between control and learning, but so far proofs of performance and safety are mostly restricted to limited cases. Thus, learning elements are often used as black boxes in the loop, with limited interpretability and less than completely understood properties. Further progress hinges on the development of a principled understanding of the limitations of control-oriented machine learning. This talk will present some initial results unveiling the fundamental limitations of some popular learning algorithms and architectures when used to control a dynamical system. For instance, it shows that even though feed forward neural nets are universal approximators, they are unable to stabilize some simple systems. We also show that a recurrent neural net with differentiable activation functions that stabilizes a non-strongly stabilizable system must itself be open loop unstable, and discuss the implications of this for training with noisy, finite data. Finally, we present a simple system where any controller based on unconstrained optimization of the parameters of a given structure fails to render the closed loop system input-to-state stable. The talk finishes by arguing that when the goal is to learn stabilizing controllers, the loss function should reflect closed loop performance, which can be accomplished using gap-metric motivated loss functions, and presenting initial steps towards that goal.
Posted August 18, 2023
Last modified September 11, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Cristina Pignotti, Università degli Studi dell'Aquila
Consensus Results for Hegselmann-Krause Type Models with Time Delay
We study Hegselmann-Krause (HK) opinion formation models in the presence of time delay effects. The influence coefficients among the agents are nonnegative, as usual, but they can also degenerate. This includes, e.g., the case of on-off influence, namely the agents do not communicate over some time intervals. We give sufficient conditions ensuring that consensus is achieved for all initial configurations. Moreover, we analyze the continuity type equation obtained as the mean-field limit of the particle model when the number of agents goes to infinity. Finally, we analyze a control problem for a delayed HK model with leadership and design a simple control strategy steering all agents to any fixed target opinion.
Posted September 12, 2023
Last modified October 11, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Melvin Leok, University of California, San Diego
Connections Between Geometric Mechanics, Information Geometry, Accelerated Optimization and Machine Learning
Geometric mechanics describes Lagrangian and Hamiltonian mechanics geometrically, and information geometry formulates statistical estimation, inference, and machine learning in terms of geometry. A divergence function is an asymmetric distance between two probability densities that induces differential geometric structures and yields efficient machine learning algorithms that minimize the duality gap. The connection between information geometry and geometric mechanics will yield a unified treatment of machine learning and structure-preserving discretizations. In particular, the divergence function of information geometry can be viewed as a discrete Lagrangian, which is a generating function of a symplectic map, that arise in discrete variational mechanics. This identification allows the methods of backward error analysis to be applied, and the symplectic map generated by a divergence function can be associated with the exact time-h flow map of a Hamiltonian system on the space of probability distributions. We will also discuss how time-adaptive Hamiltonian variational integrators can be used to discretize the Bregman Hamiltonian, whose flow generalizes the differential equation that describes the dynamics of the Nesterov accelerated gradient descent method.
Posted August 22, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Eduardo Cerpa, Pontificia Universidad Católica de Chile
SIAM Activity Group on Control and Systems Theory Prize Recipient
Control and System Theory Methods in Neurostimulation
Electrical stimulation therapies are used to treat the symptoms of a variety of nervous system disorders. Recently, the use of high frequency signals has received increased attention due to its varied effects on tissues and cells. In this talk, we will see how some methods from Control and System Theory can be useful to address relevant questions in this framework when the FitzHugh-Nagumo model of a neuron is considered. Here, the stimulation is through the source term of an ODE and the level of neuron activation is associated with the existence of action potentials which are solutions with a particular profile. A first question concerns the effectiveness of a recent technique called interferential currents, which combines two signals of similar kilohertz frequencies intended to activate deeply positioned cells. The second question is about how to avoid the onset of undesirable action potentials originated when signals that produce conduction block are turned on. We will show theoretical and computational results based on methods such as averaging, Lyapunov analysis, quasi-static steering, and others.
Posted August 22, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Philip E. Paré, Purdue University
Modeling, Estimation, and Analysis of Epidemics over Networks
We present and analyze mathematical models for network-dependent spread. We use the analysis to validate a SIS (susceptible-infected-susceptible) model employing John Snow’s classical work on cholera epidemics in London in the 1850’s. Given the demonstrated validity of the model, we discuss control strategies for mitigating spread, and formulate a tractable antidote administration problem that significantly reduces spread. Then we formulate a parameter estimation problem for an SIR (susceptible-infected-recovered) networked model, where costs are incurred by measuring different nodes' states and the goal is to minimize the total cost spent on collecting measurements or to optimize the parameter estimates while remaining within a measurement budget. We show that these problems are NP-hard to solve in general and propose approximation algorithms with performance guarantees. We conclude by discussing an ongoing project where we are developing online parameter estimation techniques for noisy data and time-varying epidemics.
Posted January 18, 2023
Last modified October 30, 2023
Control and Optimization Seminar Questions or comments?
10:30 am 233 Lockett and Zoom (Click “Questions or Comments?” to request a Zoom link)
Maruthi Akella, University of Texas
Fellow of AIAA, IEEE, and AAS
Sub-Modularity Measures for Learning and Robust Perception in Aerospace Autonomy
Onboard learning and robust perception can be generally viewed to characterize autonomy as overarching system-level properties. The complex interplay between autonomy and onboard decision support systems introduces new vulnerabilities that are extremely hard to predict with most existing guidance and control tools. In this seminar, we review some recent advances in learning-oriented and information-aware path- planning, and sub-modularity metrics for non-myopic sensor scheduling for “plug-and- play” systems. The concept of “learning-oriented” path-planning is realized through certain new classes of exploration inducing distance metrics. These technical foundations will be highlighted through aerospace applications with active learning inside dynamic and uncertain environments.
Posted September 2, 2023
Last modified November 15, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Sean Meyn, University of Florida
Robert C. Pittman Eminent Scholar Chair, IEEE Fellow, IEEE CSS Distinguished Lecturer
Stochastic Approximation and Extremum Seeking Control
Stochastic approximation was introduced in the 1950s to solve root finding problems, of which optimization is a canonical application. It is argued in recent work that extremum seeking control (ESC), a particular approach to gradient-free optimization with an even longer history, can be cast as quasi-stochastic approximation (QSA). In this lecture, we will go through the basics of these (until now) disparate fields. Application of QSA theory to ESC leads to several significant conclusions, including that ESC is not globally stable, as examples show. Careful application of QSA theory leads to new algorithms that are stable without any loss of performance. Also, QSA theory immediately provides asymptotic and transient bounds, providing guidelines for algorithm design. In addition to surveying this general theory, the talk provides a tutorial on design principles through numerical studies.
Posted September 29, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Hélène Frankowska, Sorbonne University
Differential Inclusions on Wasserstein Spaces
Optimal control in Wasserstein spaces addresses control of systems with large numbers of agents. It is well known that for optimal control of ODEs, the differential inclusions theory provides useful tools to investigate existence of optimal controls, necessary optimality conditions and Hamilton-Jacobi- Bellman equations. Recently, many models arising in social sciences used the framework of Wasserstein spaces, i.e. metric spaces of Borel probability measures endowed with the Wasserstein metric. This talk is devoted to a recent extension given in [1] of the theory of differential inclusions to the setting of general Wasserstein spaces. In the second part of the talk, necessary and sufficient conditions for the existence of solutions to state-constrained continuity inclusions in Wasserstein spaces, whose right-hand sides may be discontinuous in time, are provided; see [2]. These latter results are based on a fine investigation of the infinitesimal behavior of the underlying reachable sets, which heuristically amounts to showing that up to a negligible set, every admissible velocity can be realized as the metric derivative of a solution of the continuity inclusion, and vice versa. Building on these results, necessary and sufficient geometric conditions for the viability and invariance of stationary and time-dependent constraints, which involve a suitable notion of contingent cones in Wasserstein spaces, are established. Viability and invariance theorems in a more restrictive framework were already applied in [5], [6] to investigate stability of controlled continuity equations and uniqueness of solutions to HJB equations. The provided new tools allow us to get similar results in general Wasserstein spaces. References: [1] BONNET B. and FRANKOWSKA H., Caratheodory Theory and a Priori Estimates for Continuity Inclusions in the Space of Probability Measures, preprint https://arxiv.org/pdf/2302.00963.pdf, 2023. [2] BONNET B. and FRANKOWSKA H., On the Viability and Invariance of Proper Sets under Continuity Inclusions in Wasserstein Spaces, SIAM Journal on Mathematical Analysis, to appear. [3] BONNET B. and FRANKOWSKA H., Differential inclusions in Wasserstein spaces: the Cauchy-Lipschitz framework, Journal of Diff. Eqs. 271: 594-637, 2021. [4] BONNET B. and FRANKOWSKA H., Mean-field optimal control of continuity equations and differential inclusions, Proceedings of 59th IEEE Conference on Decision and Control, Republic of Korea, December 8-11, 2020: 470-475, 2020. [5] BONNET B. and FRANKOWSKA H., Viability and exponentially stable trajectories for differential inclusions in Wasserstein spaces, Proceedings of 61st IEEE Conference on Decision and Control, Mexico, December 6-9, 2022: 5086-5091, 2022. [6] BADREDDINE Z. and FRANKOWSKA H., Solutions to Hamilton-Jacobi equation on a Wasserstein space, Calculus of Variations and PDEs 81: 9, 2022.
Posted September 8, 2023
Last modified November 14, 2023
Control and Optimization Seminar Questions or comments?
10:30 am – 11:20 am Zoom (Click “Questions or Comments?” to request a Zoom link)
Meeko Oishi, University of New Mexico
NSF BRITE Fellow
Human-Centered Probabilistic Planning and Control
Although human interaction with autonomous systems is becoming ubiquitous, few tools exist for planning and control of autonomous systems that account for human uncertainty and decision making. We seek methods for probabilistic verification and control that can help ensure compatibility of autonomous systems with human decision making and human uncertainty. This requires the development of theory and computational tools that can accommodate arbitrary, non-Gaussian uncertainty for both probabilistic verification and control, potentially without high confidence models. This talk will focus on our work in probabilistic verification of ReLU neural nets, data-driven stochastic optimal control and stochastic reachability. Our approaches to probabilistic verification are based in Fourier transforms and chance constrained optimization, and our approaches to data-driven stochastic planning and control are based in conditional distribution embeddings. Both of these approaches enable computation without gridding, sampling, or recursion. We also present recent work on data-driven tools for high fidelity modeling and characterization of human-in-the-loop trajectories, that accommodate dynamic processes with probabilistic human inputs.