(The talk will be given in English)

Speaker:     Prof. Marc Teboulle
                     School of Mathematical Sciences, TAU

Monday, May 27^th, 2019
15:00 - 16:00

Room 011, Kitot Bldg., Faculty of Engineering

Everything Old is New Again: The Return of Gradient Based Optimization Methods
 

Abstract

The gradient method, forged by Cauchy about 170 years ago, has regained over the last decade a strong revived interest in modern optimization through many of its variants and relatives known as First Order Methods (FOM). This renewed interest in  FOM has emerged  from  the current high demand in solving optimization problems arising in a wide spectrum of modern applications,  e.g., in signal processing, image science, machine learning, and physics. These applied problems are often ill-posed, nonsmooth,  convex or nonconvex,  and typically very large or even huge scale. This rules out the option of using sophisticated algorithms (e.g., Newton types schemes involving inversion of matrices) which often become prohibitively expensive.  Elementary first order methods using function values and gradient/subgradient information then often remain our best alternative to tackle such large scale optimization problems.  In turn, this rich collection of applied problems is providing fresh perspectives for optimization algorithms leading to new fundamental research with challenging theoretical and computational questions in the field.

We discuss recent advances on the design and analysis of gradient-based algorithms that we have developed and applied successfully in various areas.   In particular, we highlight the ways in which mathematical structures and data information can be beneficially exploited to design and analyze simple convex and nonconvex optimization methods. The talk is intended to a wide audience, and we will assume (almost) no prior knowledge in continuous optimization.

Short Bio
Marc Teboulle is the Eric and Sheila Samson Chair and Professor of Optimization at the School of Mathematical Sciences of Tel Aviv University. He received his D.Sc. from the Technion, Israel Institute of Technology. Teboulle research interests are in the area of continuous optimization, including theoretical foundation and algorithmic analysis, as well as its applications to many areas of science and engineering. Prior to joining Tel Aviv University in 1996, Teboulle was an applied mathematician for the Israel Aircraft Industries, a post-doctoral fellow at Dalhousie University, Canada, and a Professor of Mathematics at University of Maryland (USA). 
Teboulle has published widely on optimization theory, algorithms, and applications.  He is a co-author of the book "Asymptotic Cones and Functions in Optimization and Variational Inequalities" (Springer Monographs in Mathematics, 2003, with A. Auslender). He is a Fellow of SIAM, the Society for Industrial and Applied Mathematics. He served a term as Area Editor for Continuous Optimization of Mathematics of Operations Research from 2013-2018. He currently serves as Corresponding Editor for the Optimization Area of COCV-- Control, Optimization, and Calculus of Variations,  and is on the editorial boards of some other leading journals, including SIAM Journal of Optimization, and SIAM Journal on Mathematics of Data Science.‏

הקמת פרויקטי תשתית תוך יצירת דו שיח סביבתי

מקרה בוחן: הולכת הגז הטבעי בישראל

יום רביעי 12.06.19 התכנסות:  18:30-19:15 הרצאה: 19:15-20:30

בפקולטה להנדסה אוניברסיטת תל אביב

מרצה: רחל לוטן מנהלת המחלקה לתכנון סטטוטורי, הסדרת מקרקעין ופיתוח בר קיימא ב"נתיבי הגז הטבעי לישראל בע"מ". 

לפרטים נוספים ולהרשמה לחצו כאן 

School of Mechanical Engineering Seminar
Monday, June 3, 2019 at 14:00
Wolfson Building of Mechanical Engineering, Room 206

Data-Driven Modeling of Pitching and Plunging Wings in Single and Tandem Configurations in Hovering Flight

Victor Troshin

Academic Advisor:

Prof. Avi Seifert

The presented work addresses the challenges related to the low order dynamic modeling of the fluid domain with immersed moving solid bodies in it. In this work mathematical tools were developed and tested on numerical and experimental data. The ultimate goal of this work was to develop a methodology which enables a real-time flow field estimation based on a small number of sensors. As an example of such a problem, modeling of pitching and plunging wings in single and tandem configurations was chosen. In order to achieve the defined goal, the presented research was carried out in the three following stages:

In the first stage of the research, a proper orthogonal decomposition (POD) methodology for a flow field in a domain with moving boundaries was developed. In the standard POD approach, the properties of the region of the domain which alternatingly occupied by fluid and solid are not defined. Thus, here, prior to the decomposition, the domain with moving or deforming boundaries was mapped to a stationary domain using volume preserving mapping. This mapping was created by combining a transfinite interpolation and volume adjustment algorithm. The algorithm is based on an iterative solution of the Laplace equation with respect to the displacement potential of the grid points. At this stage of the research, the method was validated on CFD simulation of pitching and plunging ellipse in a still fluid.

The main goal of the second stage of the research was to develop a low order model of a heaving airfoil in a still fluid using experimental measurements. This was achieved by modifying and applying the tools developed in the first stage of the study.  The modified POD approach together with a time delay neural network (TDNN) was used to model and predict the flow field evolution using only a couple of low profile load sensors. The neural network estimated the amplitudes of the most energetic modes using four sensory inputs. The modes were calculated using the proper orthogonal decomposition of the flow field data obtained experimentally by time-resolved, phase-locked particle imaging velocimetry (TRPIV). The model showed good estimation quality.

In the final stage of this study, a modeling methodology was implemented on measured flow field data of pitching and plunging wings in a tandem configuration. Here, the velocity field associated with the wings’ flapping motion was mapped and modeled using the previously developed POD approach. The flow field dynamics was approximated by a linear model based on only four POD modes. Since the state of the low order model (i.e. the amplitudes of the modes) is physically impossible to measure, a Kalman filter was implemented. The Kalman filter used the signals from two low-profile strain gauge sensors located at the root of the hindwing to evaluate the reduced-order state of the system. Then, the full state of the system was estimated using the POD approximation. Therefore, by using only two strain sensors’ signals, the complex vortex dynamics associated with the tandem wings motion was successfully modeled.

Speaker: Daniel Brodeski

M.Sc. student under the supervision of Dr. Raja Giryes

Wednesday, June 19^th, 2019 at 15:30

Room 011, Kitot Bldg., Faculty of Engineering

Deep Radar Detector

Abstract

While camera and LiDAR processing have been revolutionized since the introduction of deep learning, radar processing still relies on classical tools. In this work, we introduce a deep learning approach for radar processing, working directly with the radar complex data. To overcome the lack of radar labeled data, we rely in training only on the radar calibration data and introduce new radar augmentation techniques. We evaluate our method on the radar 4D detection task and demonstrate superior performance compared to the classical approaches while keeping real-time performance. Applying deep learning on radar data has several advantages such as eliminating the need for an expensive radar calibration process each time and enabling classification of the detected objects with almost zero-overhead.