Updated on 2025/04/02

Information

 

写真a

 
THEMELIS ANDREAS
 
Organization
Faculty of Information Science and Electrical Engineering Department of Electrical Engineering Associate Professor
School of Engineering Department of Electrical Engineering and Computer Science(Concurrent)
Graduate School of Information Science and Electrical Engineering Department of Electrical and Electronic Engineering(Concurrent)
Joint Graduate School of Mathematics for Innovation (Concurrent)
Title
Associate Professor
Contact information
メールアドレス
Tel
0928023718
Profile
The increase of computing capability and the development of powerful (micro)processors that we have witnessed in the last years has motivated engineers to design more sophisticated ad-hoc control strategies based on "Optimization". The importance of this science is dictated by the fact that virtually any engineering problem is (or can be reduced to) a functional minimization. For instance, finding the "best" route in path planning amounts to finding the one that minimizes a cost (function), which is the contribution of factors such as distance, time, fuel consumption, and so on. The main challenge is to find a suitable balance between convergence speed, low computational requirements, and range of problems that can be solved. Our group aims at developing efficient algorithms to be employed in a wide area of engineering applications, including, but not limited to, control and signal processing.
External link

Degree

  • PhD

Research History

  • KU Leuven (Postdoctoral researcher)   

Research Interests・Research Keywords

  • Research theme: The increase of computing capability and the development of powerful (micro)processors that we have witnessed in the last years has motivated engineers to design more sophisticated ad-hoc control strategies based on "Optimization". The importance of this science is dictated by the fact that virtually any engineering problem is (or can be reduced to) a functional minimization. For instance, finding the "best" route in path planning amounts to finding the one that minimizes a cost (function), which is the contribution of factors such as distance, time, fuel consumption, and so on. The main challenge is to find a suitable balance between convergence speed, low computational requirements, and range of problems that can be solved. Our group aims at developing efficient algorithms to be employed in a wide area of engineering applications, including, but not limited to, control and signal processing.

    Keyword: Optimization algorithms for engineering

    Research period: 2021.1 - 2021.6

Papers

▼display all

Presentations

  • Adaptive proximal gradient methods for convex bilevel optimization Invited International conference

    Andreas Themelis, Puya Latafat, Silvia Villa, Panagiotis Patrinos

    Control & Optimisation (ContrOpt) Pisa 2023  2023.5 

     More details

    Event date: 2023.5

    Language:English   Presentation type:Oral presentation (general)  

    Venue:University of Pisa   Country:Italy  

    Bilevel optimization is a comprehensive framework that bridges single- and multi-objective optimization. It encompassess many general formulations, such as, but not limited to, standard nonlinear programs. This work demonstrates how elementary proximal gradient iterations can be used to solve a wide class of convex bilevel optimization problems without involving subroutines. Compared to and improving upon existing methods, ours (1) can handle a much wider class of problems, including both constraints and nonsmooth terms, (2) does not require strong convexity or Lipschitz smoothness assumptions, and (3) provides a systematic adaptive stepsize selection strategy with no need of function evaluations. A linesearch-free variant is also proposed that eliminates wasteful backtracking trials at the sole expense of cost evaluations.

    Other Link: https://contropt2023.ec.unipi.it

    Repository Public URL: https://hdl.handle.net/2324/6790346

  • Inertia and relative smoothness in nonconvex minimization: a case study on the forward-reflected-backward algorithm Invited International conference

    Andreas Themelis, @Ziyuan Wang, #Hongjia Ou, Xianfu Wang

    2022 International Workshop on Continuous Optimization  2022.12 

     More details

    Event date: 2022.12

    Language:English   Presentation type:Symposium, workshop panel (public)  

    Venue:Tokyo Institute of Technology   Country:Japan  

    The use of momentum to accelerate convergence of first-order algorithms has been gaining renewed interest ever since its first appearance 60 years back. Initially inspired by the physics intuition that inertia is effective in preventing oscillatory behaviors, momentum-type techniques are typically designed in attempt to improve convergence speed. While this effect can only be achieved and justified for positive momentum coefficients, our study on the forward-reflected-backward splitting of Malitsky and Tam suggests the necessity of "negative" values to guarantee convergence under mere relative smoothness assumptions, for nonconvex problems. Our conclusions are in line with, and more pessimistic than, a similar conjecture of Dragomir et al. for the mirror descent algorithm.

    Other Link: http://www.opt.c.titech.ac.jp/DecemberWorkshop/

    Repository Public URL: https://hdl.handle.net/2324/6790347

  • Splitting algorithms for nonconvex optimization: unified analysis and Newton-type acceleration Invited International conference

    Andreas Themelis

    Northwestern Polytechnical University Optimization Seminar  2022.10 

     More details

    Event date: 2022.10

    Language:English   Presentation type:Public lecture, seminar, tutorial, course, or other speech  

    Country:China  

    We provide a unified interpretation of splitting algorithms for nonconvex optimization through the lens of majorization-minimization. Possibly under assumptions to compensate the lack of convexity, this setting is general enough to cover ADMM as well as forward-backward, Douglas-Rachford and Davis-Yin splittings. Proximal envelopes, a generalization of the Moreau envelope, are shown to be natural merit functions for establishing convergence results. Their regularity properties also enable the integration of fast direction of quasi-Newton-type, that differently from any other approach for nonsmooth optimization preserve the same operation complexity of the original splitting scheme.

    Repository Public URL: https://hdl.handle.net/2324/6790348

  • Efficient lightweight solvers for real-time embedded nonlinear MPC Invited

    Andreas Themelis

    60th SICE Annual Conference 2021  2021.9 

     More details

    Event date: 2021.9

    Language:English   Presentation type:Oral presentation (general)  

    Country:Japan  

    Model predictive control (MPC) has become a popular strategy to implement feedback control loops for a variety of systems. Since most systems are nonlinear by nature, nonlinear MPC offers a more accurate modeling, but it leads to nonconvex and much more complicated problems that need to be solved at every sampling step. In "embedded" applications such as autonomous driving, the resulting problems easily become of large scale and the sampling time can be as low as few milliseconds, thus imposing an imperative demand for algorithmic speed and efficiency.
    In this talk we show how the scalability properties of "proximal algorithms" can conveniently be employed to design certifiable, fast, and lightweight algorithms perfectly suited for embedded applications.

    Other Link: https://www.imi.kyushu-u.ac.jp/wp-content/uploads/2022/07/mil_84.pdf

  • Optimization for real-time control with limited resources Invited International conference

    Andreas Themelis

    6th IFAC Conference on Engine and Powertrain Control, Simulation and Modeling  2021.8 

     More details

    Event date: 2021.8

    Language:English   Presentation type:Public lecture, seminar, tutorial, course, or other speech  

    Country:Japan  

    Model predictive control (MPC) has become a popular strategy to implement feedback control loops for a variety of systems. An MPC strategy aims at repeatedly selecting control inputs that yield the best outcome among all possible choices. To assess the quality of an input, a cost function is designed that takes into account the desired goals, such as going from point A to point B in short time without wasting fuel. This leads to a problem formulation where the objective is the "minimization" of a cost function, encoding the desired goal, subject to constraints, which instead account for actuators limitations (e.g. maximum speed or power) as well as environmental impediments such as physical obstacles or speed limits.
    Most systems in nature and science evolve according to "nonlinear" laws, and this leads to the major challenge of nonsmooth and nonconvex problems that need to be solved within sampling time, that is, before a new control input needs to be fed again to the system. In "embedded" applications such as autonomous driving, the resulting problems easily become of large scale and the sampling time can be as low as few milliseconds, thus imposing an imperative demand for algorithmic speed and efficiency.
    In this talk we show how the scalability properties of "proximal algorithms" can conveniently be employed to design certifiable, fast, and lightweight algorithms perfectly suited for embedded applications.

    Other Link: https://ifac.papercept.net/conferences/conferences/ECOSM21/program/ECOSM21_ContentListWeb_1.html#suts1

    Repository Public URL: https://hdl.handle.net/2324/6790350

  • Bregman proximal algorithms for composite and finite-sum nonconvex minimization problems Invited International conference

    Andreas Themelis, Puya Latafat, Masoud Ahookhosh, Panagiotis Patrinos

    SIAM Conference on Optimization (OP21)  2021.7 

     More details

    Event date: 2021.7

    Language:English   Presentation type:Symposium, workshop panel (public)  

    Venue:[online]  

    The employment of Bregman divergences in splitting algorithms has been growing in popularity in the last few years. Firstly, the extra degree of freedom in the metric selection can lead to new algorithms or may provide new insights on known ones. Secondly, while many classical such schemes are bound to Lipschitz differentiability requirements (especially in the nonconvex setting), the recently introduced notion of ``relative smoothness'' has considerably widened the range of problems that can be addressed.

    This talk deals with Bregman proximal algorithms in the fully nonconvex setting. The employment of the Bregman Moreau envelope as Lyapunov function leads to extremely simple and intuitive converge analyses that naturally extend to block-coordinate variants. Furthermore, continuity of the envelope allows one to design linesearch-type extensions that preserve oracle complexity and convergence properties of first-order (Bregman) splitting schemes, and yet can attain up to superlinear asymptotic rates when directions of quasi-Newton type are selected.

    Other Link: https://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=72026

    Repository Public URL: https://hdl.handle.net/2324/6790349

  • A universal majorization-minimization framework for the convergence analysis of nonconvex proximal algorithms Invited International conference

    Andreas Themelis, Panagiotis Patrinos

    6th International Conference on Continuous Optimization  2019.8 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Public lecture, seminar, tutorial, course, or other speech  

    Venue:Berlin   Country:Germany  

  • Proximal envelopes Invited International conference

    Andreas Themelis, Panagiotis Patrinos

    17th IEEE European Control Conference  2018.6 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Public lecture, seminar, tutorial, course, or other speech  

    Venue:Limassol   Country:Cyprus  

  • A simple and efficient algorithm for Nonlinear MPC International conference

    Lorenzo Stella, Andreas Themelis, Pantelis Sopasakis, Panagiotis Patrinos

    56th IEEE Conference on Decision and Control  2017.12 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Oral presentation (general)  

    Venue:Melbourne   Country:Australia  

  • Accelerated Douglas-Rachford splitting and ADMM for structured nonconvex optimization Invited International conference

    Panagiotis Patrinos, Andreas Themelis

    CMO-BIRS Workshop on Splitting Algorithms, Modern Operator Theory and Applications  2017.9 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Public lecture, seminar, tutorial, course, or other speech  

    Venue:Oaxaca   Country:Mexico  

  • Newton-type proximal algorithms for nonconvex optimization Invited International conference

    Andreas Themelis

    LCCC focus period on large scale and distributed optimization  2017.6 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Public lecture, seminar, tutorial, course, or other speech  

    Venue:Lund   Country:Sweden  

  • Newton-type operator splitting algorithms International conference

    Andreas Themelis, Puya Latafat, Panagiotis Patrinos

    4th European Conference on Computational Optimization  2016.9 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Oral presentation (general)  

    Venue:Leuven   Country:Belgium  

  • A variable metric Stochastic Gradient method for large scale optimization International conference

    Andreas Themelis, Silvia Villa, Panagiotis Patrinos, Alberto Bemporad

    28th European Conference on Operational Research  2016.7 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Oral presentation (general)  

    Venue:Poznan   Country:Poland  

  • Stochastic Gradient Methods for Stochastic Model Predictive Control Invited International conference

    Andreas Themelis, Silvia Villa, Alberto Bemporad, Panagiotis Patrinos

    15th IEEE European Control Conference  2016.6 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Oral presentation (general)  

    Venue:Aalborg   Country:Denmark  

  • A globally and superlinearly convergent algorithm for finding fixed points of nonexpansive operators Invited International conference

    Andreas Themelis, Puya Latafat, Panagiotis Patrinos

    50th Anniversary of the Center for Operations Research and Econometrics  2016.5 

     More details

    Event date: 2021.6

    Language:English   Presentation type:Oral presentation (general)  

    Venue:Louvain la Neuve   Country:Belgium  

▼display all

Academic Activities

  • Screening of academic papers

    Role(s): Peer review

    2022

     More details

    Type:Peer review 

    Number of peer-reviewed articles in foreign language journals:6

    Proceedings of International Conference Number of peer-reviewed papers:2

  • Screening of academic papers

    Role(s): Peer review

    2021

     More details

    Type:Peer review 

    Number of peer-reviewed articles in foreign language journals:7

  • Screening of academic papers

    Role(s): Peer review

    2020

     More details

    Type:Peer review 

    Number of peer-reviewed articles in foreign language journals:5

    Proceedings of International Conference Number of peer-reviewed papers:3

  • Screening of academic papers

    Role(s): Peer review

    2019

     More details

    Type:Peer review 

    Number of peer-reviewed articles in foreign language journals:2

  • Screening of academic papers

    Role(s): Peer review

    2018

     More details

    Type:Peer review 

    Number of peer-reviewed articles in foreign language journals:1

  • Screening of academic papers

    Role(s): Peer review

    2017

     More details

    Type:Peer review 

    Number of peer-reviewed articles in foreign language journals:1

    Proceedings of International Conference Number of peer-reviewed papers:1

  • Screening of academic papers

    Role(s): Peer review

    2016

     More details

    Type:Peer review 

    Proceedings of International Conference Number of peer-reviewed papers:1

  • Screening of academic papers

    Role(s): Peer review

    2015

     More details

    Type:Peer review 

    Proceedings of International Conference Number of peer-reviewed papers:2

▼display all

Research Projects

  • The main objective of the project is to develop learning-based techniques for devising ad-hoc tuning-free optimization algorithms for convex and nonconvex optimization problems. A novel universal framework will be developed, which will serve as a solid theoretical ground for the development of new learning paradigms to train optimization methods subject to certificates of (speed of) conver-gence and quality of output solution.

    2021.6

    Joint research

      More details

    Authorship:Coinvestigator(s)  Grant type:Other funds from industry-academia collaboration

  • The project is concerned with optimization algorithms for engineering, in compliance with the application challenges: efficiency, limited power of microprocessors, and nonconvexity of the problems. The final goal is to provide efficient open‑source multi‑purpose software with theoretical guarantees.

    Grant number:21K17710  2021 - 2023

    Japan Society for the Promotion of Science  Grants-in-Aid for Scientific Research  Early-Career Scientists

      More details

    Authorship:Principal investigator  Grant type:Scientific research funding

Class subject

  • プログラミング演習Ⅲ

    2023.12 - 2024.2   Winter quarter

  • 電気電子工学研究調査(第3グループ)

    2023.10 - 2024.3   Second semester

  • Presentations in EEE Studies(Group3)

    2023.10 - 2024.3   Second semester

  • (IUPE)Mathematics for EECS

    2023.4 - 2023.6   Spring quarter

  • プログラミング演習Ⅲ

    2022.12 - 2023.2   Winter quarter

  • (IUPE)Mathematics for EECS

    2022.4 - 2022.6   Spring quarter

  • (IUPE)Mathematics for EECS

    2021.4 - 2021.6   Spring quarter

▼display all

Travel Abroad

  • 2023.8

    Staying countory name 1:Japan   Staying institution name 1:10th International Congress on Industrial and Applied Mathematics

  • 2023.8

    Staying countory name 1:China   Staying institution name 1:Chongqing Normal University

  • 2023.5

    Staying countory name 1:Italy   Staying institution name 1:University of Pisa

  • 2022.11 - 2022.12

    Staying countory name 1:Japan   Staying institution name 1:Tokyo Institute of Technology

  • 2019.6

    Staying countory name 1:Italy   Staying institution name 1:IMT School for Advanced Studies Lucca

  • 2018.8

    Staying countory name 1:Germany   Staying institution name 1:6th International Conference on Continuous Optimization

  • 2018.6

    Staying countory name 1:Cyprus   Staying institution name 1:17th IEEE European Control Conference

  • 2017.12

    Staying countory name 1:Australia   Staying institution name 1:56th IEEE International Conference on Decision and Control

  • 2017.6

    Staying countory name 1:Sweden   Staying institution name 1:Department of Automatic Control

  • 2016.7

    Staying countory name 1:Poland   Staying institution name 1:28th European Conference on Operational Research

  • 2016.6

    Staying countory name 1:Denmark   Staying institution name 1:15th IEEE European Control Conference

  • 2015.10 - 2016.4

    Staying countory name 1:Belgium   Staying institution name 1:KU Leuven

▼display all