Luis A. Aguirre is a full professor at the Department of Electronics Engineering at UFMG, Brazil. His research includes the identification of nonlinear system, nonlinear dynamics, control systems, synchronization and analysis of dynamical networks.
Full BIO
Luis A. Aguirre received a PhD degree in 1994 from the University of Sheffield, England and is a full professor at the Department of Electronics Engineering at UFMG. He is the author of four books and was the Editor-in-Chief of Enciclopédia de Automática (3 volume set), sponsored by the Brazilian Society of Automation (SBA) and published by Editora Blücher. From 2009 to 2012 he served as the Editor-in-Chief of Controle & Automação: Revista da Sociedade Brasileira de Automática, currently published by Springer Verlag under the name Journal of Control, Automation and Electrical Systems. His research includes the identification of nonlinear system, nonlinear dynamics, control systems, synchronization and analysis of dynamical networks.
Talk: Taking Advantage of Auxiliary Information in Nonlinear System Identification
Classical system identification consists of building a dynamical mathematical model from a set of dynamical data, usually collected during a carefully designed test. One of the challenges nowadays is to take advantage of existing data that are recorded during operation. The good side of this is that such data usually abound, but the other side of the coin is less attractive: whereas the amount of data is great, the underlying dynamical information is not always what is needed to build a good model. One way of overcoming this situation is to, taking advantage of the data, dig out specific features from the process. In many problems in the field of control engineering both dynamical and steady-state information are critical and should be used to build a competitive model. Unfortunately it often happens that one set of data does not contain all that is needed, as assumed in classical system identification. This talk will address the problem of using and therefore benefitting from auxiliary information in nonlinear system identification. For the sake of discussion and illustration the scenario where dynamical data do not contain sufficient steady-state information will be addressed. In this case the static behavior of the process will have to be estimated from other windows of data and will compose the auxiliary information. Ways to use such information simultaneously with the dynamical set of data will be discussed and illustrated using data from an oil well.
Roy Smith is a Professor in the Automatic Control Laboratory at the Swiss Federal Institute of Technology, ETH Zurich in Switzerland. His research interests include: the identification and control of uncertain systems, and distributed estimation, communication and control systems.
Full BIO
Roy Smith is a Professor in the Automatic Control Laboratory at the Swiss Federal Institute of Technology (ETH, Zurich) in Switzerland. From 1990 to 2010 he was on the faculty of the Electrical Engineering Dept. at the University of California, Santa Barbara. He received his undergraduate education at Canterbury University in New Zealand (1980) and a Ph.D. from the California Institute of Technology (1990). His research interests include: the identification and control of uncertain systems, and distributed estimation, communication and control systems. His application experience includes: process control, automotive engines, flexible space structures, aeromanoeuvring Mars entry vehicles, formation flying of spacecraft, magnetically levitated bearings, high energy accelerator control, thermoacoustic engines, airborne wind energy, and energy control for buildings. He has been a long time consultant to the NASA Jet Propulsion Laboratory on guidance, navigation and control aspects of interplanetary and deep space science spacecraft. He is a Fellow of the IEEE & IFAC, an Associate Fellow of the AIAA, and a member of SIAM and NZAC.
Mohammad Khosravi is a doctoral student at Automatic Control Laboratory, ETH Zurich in Switzerland. His research interest includes data-driven modeling and decision-making, and their applications in buildings, energy, and industrial systems.
Full BIO
Mohammad Khosravi received a B.Sc. in electrical engineering and a B.Sc. in mathematical sciences from Sharif University of Technology, Tehran, Iran, in 2011. He obtained a postgraduate diploma in mathematics from ICTP, Trieste, Italy, in 2012. He was a research assistant in the mathematical biology group at Institute for Research in Fundamental Sciences, Tehran, Iran, in 2012-2014. He received his M.A.Sc. degree in electrical and computer engineering from Concordia University, Montreal, Canada, in 2016. Currently, he is a Ph.D. candidate at Automatic Control Laboratory, ETH Zurich. He has won several awards, including the gold medal of the National Mathematics Olympiad, the Outstanding Student Paper Award in CDC 2020, and the Outstanding Reviewer Award for IEEE Journal of Control Systems Letters. His research interests are on data-driven and learning-based methods in modeling, control & optimization, and their applications in buildings, energy, and industrial systems.
Talk: Identification with side information: exploiting information about the region of convergence (joint talk of Roy Smith & Mohammad Khosravi)
Dynamical systems are omnipresent in various fields of science and different modern technologies. Being in the era of data, one may pose the question of how to identify or learn a dynamical system using collected data. In many situations, in addition to data, we may have some side information about the dynamics to be incorporated in the identification problem in a correct and tractable manner. The origin of this side information can be the physics, the nature of the system or it may come from observed behaviours in data or past experiments. We present a method for the case where the side information is knowledge about the region of attraction (ROA) of an equilibrium point, i.e., a subset of the ROA containing the equilibrium point is known. The identification method is based on an optimisation problem minimising the fitting error while guaranteeing the incorporation of the knowledge about the ROA. The side information is integrated into the identification problem by considering an unknown Lyapunov function verifying the stability property. The resulting problem is a non-convex infinite-dimensional optimisation with an infinite number of constraints. To obtain a tractable formulation, a suitably designed finite subset of the constraints is considered. The subsequent problem admits a solution in form of a linear combination of the feature maps of the kernel and their derivatives. This linear parametric form allows us to formulate an equivalent finite-dimensional optimisation problem with a quadratic cost function subject to linear and bilinear constraints. We demonstrate the method by several examples and verify the advantage of incorporation of side information in this identification problem.
Rose Yu is an assistant professor at the University of California San Diego, Department of Computer Science and Engineering. A particular emphasis of her research is on physics-guided AI which aims to integrate first-principles with data-driven models.
Full BIO
Rose Yu is an assistant professor at the University of California San Diego, Department of Computer Science and Engineering. She was a Postdoctoral Fellow at the California Institute of Technology. Her research focuses on advancing machine learning techniques for large-scale spatiotemporal data analysis, with applications to sustainability, health, and physical sciences. A particular emphasis of her research is on physics-guided AI which aims to integrate first-principles with data-driven models. She has won Faculty Research Awards from Google, Amazon, and Adobe, several Best Paper Awards, Best Dissertation Award from USC, and was nominated as one of the "MIT Rising Stars in EECS".
Talk: Equivariant Neural Networks for Learning Spatiotemporal Dynamics
Applications such as climate science and transportation require learning complex dynamics from large-scale spatiotemporal data. Existing machine learning frameworks are still insufficient to learn spatiotemporal dynamics as they often fail to exploit the underlying physics principles. Representation theory can be used to describe and exploit the symmetry of the dynamical system. We will show how to design neural networks that are equivariant to various symmetries for learning spatiotemporal dynamics. Our methods demonstrate significant improvement in prediction accuracy, generalization, and sample efficiency in forecasting turbulent flows and predicting real-world trajectories. This is joint work with Robin Walters, Rui Wang and Jinxi Li.
Dionysios Kalogerias is an assistant professor of Electrical Engineering at Yale University. His research interests lie in the areas of machine learning, reinforcement learning, optimization, signal processing, sequential decision making, and risk.
Full BIO
Dionysios Kalogerias is an assistant professor of Electrical Engineering at Yale University. He received the PhD degree in Electrical and Computer Engineering (ECE) from Rutgers, The State University of New Jersey (2017), the MSc degree in Signal Processing and Communications from the University of Patras, Greece (2012), and the Meng degree in Computer Engineering and Informatics, also from the University of Patras, Greece (2010). Before joining Yale, Dionysios was an assistant professor with the Department of ECE, Michigan State University (2020 – 2021). Prior to that, he held postdoctoral appointments with the Department of Electrical and Systems Engineering, University of Pennsylvania (2019 – 2020), and the Department of Operations Research and Financial Engineering, Princeton University (2017 – 2019). Dionysios’ research is in the areas of machine learning, reinforcement learning, optimization, signal processing, sequential decision making, and risk, and their applications in autonomous networked systems, wireless communications, security and privacy, and system trustworthiness. He has received several awards for his work, including the Best Paper Award at ICASSP 2020, and the Best Paper of the Special Sessions Award at ICASSP 2016.
Talk: Achieving Noisy Linear Convergence in CV@R Statistical Learning: SGD, Strongly Convex Losses and Beyond
Risk-awareness is becoming an increasingly important issue in modern statistical learning theory and practice, especially due to the need to meet strict reliability requirements in high-stakes, critical applications. Examples appear naturally in many areas, such as energy, finance, robotics, radar/lidar, networking and communications, autonomy, safety, and the Internet-of-Things. In such settings, risk-aware learning formulations are particularly appealing, since they can explicitly balance the performance of optimal predictors between average-case and “difficult” to learn, infrequent, or worst-case examples, inducing a form of statistical robustness in the learning outcome. In particular, among all measures of risk, Conditional Value-at-Risk (CV@R) is by far one of the most popular (if not the most), and has been recently considered as a performance criterion in supervised statistical learning, as it is also related to desirable operational features, such as safety, fairness, distributional robustness, and prediction error stability. However, due to its variational definition, CV@R is commonly believed to result in difficult optimization problems, even for smooth and strongly convex loss functions. In this talk, we present new results disproving this statement, establishing noisy (i.e., fixed accuracy) linear convergence of stochastic gradient descent for sequential CV@R learning, for a large class of not necessarily strongly-convex (or even convex) loss functions satisfying a set-restricted Polyak-\L ojasiewicz inequality. This class contains all smooth and strongly convex losses as special cases, confirming that classical problems, such as linear least squares regression, can be solved efficiently under the CV@R criterion, just as their risk-neutral versions. We then illustrate our results numerically on indicative risk-aware learning tasks, also verifying their validity in practice.
Simone Formentin is an associate professor at Politecnico di Milano, Italy. His research interests include system identification and data-driven control with a focus on automotive and financial applications.
Full BIO
Simone Formentin was born in Legnano, Italy, in 1984. He received his B.Sc. and M.Sc. degrees cum laude in Automation and Control Engineering from Politecnico di Milano, Italy, in 2006 and 2008, respectively. In 2012, he obtained his Ph.D. degree cum laude in Information Technology within a joint program between Politecnico di Milano and Johannes Kepler University of Linz, Austria. After that, he held two postdoc appointments at EPFL, Switzerland, and the University of Bergamo, Italy. Since 2014, he has been with Politecnico di Milano, first as an assistant professor, then as an associate professor. He is the chair of the IEEE TC on System Identification and Adaptive Control and a member of the IFAC TC on Modelling, Identification and Signal Processing. He is an Associate Editor of the European Journal of Control and the IEEE CSS Conference Editorial Board. His research interests include system identification and data-driven control with a focus on automotive and financial applications.
Talk: Control-oriented regularization for dynamical system identification
In this talk, we develop a novel theoretical framework for control-oriented identification, based on a Bayesian perspective on modeling. Specifically, we show that closed-loop specifications can be incorporated within the identification procedure as a prior of the model probability distribution via suitable regularization. The corresponding kernel varies according to the additional penalty term and provides a new insight on control-oriented identification. Finally, we show that a Bayesian robust control design approach can be derived so as to exploit all the information coming from the above modeling procedure, including the estimate of the uncertainty set.
Dorsa Sadigh is an assistant professor in Computer Science and Electrical Engineering at Stanford University. Her research interests lie in the intersection of robotics, learning, and control theory.
Full BIO
Dorsa Sadigh is an assistant professor in Computer Science and Electrical Engineering at Stanford University. Her research interests lie in the intersection of robotics, learning, and control theory. Specifically, she is interested in developing algorithms for safe and adaptive human-robot interaction. Dorsa has received her doctoral degree in Electrical Engineering and Computer Sciences (EECS) from UC Berkeley in 2017, and has received her bachelor’s degree in EECS from UC Berkeley in 2012. She is awarded the NSF CAREER award, the AFOSR Young Investigator award, the IEEE TCCPS early career award, the Google Faculty Award, and the Amazon Faculty Research Award.
Erdem Bıyık is a fifth-year Ph.D. candidate in the Electrical Engineering department at Stanford. He works on developing learning algorithms for robots that actively query the humans and on training the robot policies that influence the humans or other robots to get more cooperative, which in turn improves the payoff for all agents.
Full BIO
Erdem Bıyık is a fifth-year Ph.D. candidate in the Electrical Engineering department at Stanford. He has received his B.Sc. degree from Bilkent University, Turkey, in 2017; and M.Sc. degree from Stanford University in 2019. He is interested in human-robot interaction and multi-agent systems. Specifically, he works on developing learning algorithms for robots that actively query the humans and on training the robot policies that influence the humans or other robots to get more cooperative, which in turn improves the payoff for all agents.
Talk: The Role of Representations in Human-Aware Learning and Control (joint talk of Dorsa Sadigh and Erdem Biyik)
In this talk, we will discuss a formalism for human-robot interaction built upon ideas from representation learning. Specifically, We will first discuss the notion of latent strategies — low dimensional representations sufficient for capturing non-stationary interactions. We will then talk about the challenges of learning such representations when interacting with humans, and how we can develop data-efficient techniques that enable actively learning computational models of human behavior from demonstrations, preferences, or physical corrections. Finally, we will introduce an intuitive control paradigm that enables seamless collaboration based on learned representations, and further discuss how that can be used for further influencing humans.
Andreea Bobu is a Ph.D. student at the University of California Berkeley. Her research interests lie at the intersection of machine learning, robotics, and human-robot interaction, with a focus in robot learning with uncertainty.
Full BIO
Andreea Bobu is a Ph.D. student with the Interactive Autonomy and Collaborative Technologies Laboratory (InterACT) at University of California Berkeley. She received her Bachelor’s degree in Computer Science and Engineering from MIT in 2017. Her research interests lie at the intersection of machine learning, robotics, and human-robot interaction, with a focus in robot learning with uncertainty.
Marius Wiggert is a Ph.D. student at the University of California, Berkeley. His research interests lie at the intersection of machine learning, robotics, and control theory with a focus on robot learning and applications in carbon sequestration.
Full BIO
Marius Wiggert is a Ph.D. student with the Hybrid Systems Laboratory at the University of California, Berkeley. In 2017 he received his B.Sc. in Engineering Science from the TUM in 2017 and his B.A. in Philosophy from LMU. His research interests lie at the intersection of machine learning, robotics, and control theory with a focus on robot learning and applications in carbon sequestration.
Talk: Feature Expansive Reward Learning: Rethinking Human Input (joint talk of Andreea Bobu & Marius Wiggert)
When a person is not satisfied with how a robot performs a task, they can intervene to correct it. Reward learning methods enable the robot to adapt its reward function online based on such human input, but they rely on handcrafted features. When the correction cannot be explained by these features, recent work in deep Inverse Reinforcement Learning (IRL) suggests that the robot could ask for task demonstrations and recover a reward defined over the raw state space. Our insight is that rather than implicitly learning about the missing feature(s) from demonstrations, the robot should instead ask for data that explicitly teaches it about what it is missing. We introduce a new type of human input in which the person guides the robot from states where the feature being taught is highly expressed to states where it is not. We propose an algorithm for learning the feature from the raw state space and integrating it into the reward function. By focusing the human input on the missing feature, our method decreases sample complexity and improves generalization of the learned reward over the above deep IRL baseline. We show this in experiments with a physical 7DOF robot manipulator, as well as in a user study conducted in a simulated environment.