BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:NCCR Automation Seminar: Asynchronous optimization through the len
 s of control theory
DTSTART:20240325T150000
DTEND:20240325T160000
DTSTAMP:20260407T163919Z
UID:689fbbb0244e0f9bdc3bee47cd2eb32d6e50094345769ed763e4d2af
CATEGORIES:Conferences - Seminars
DESCRIPTION:Prof. Mikael Johansson\n\nAbstract: With the ubiquitous digi
 talization of society\, decision problems are rapidly expanding in size an
 d scope. Increasingly often\, we face problems where data\, computations\,
  and decisions need to be distributed on multiple nodes. These nodes may b
 e individual cores in a CPU\, different processors in a multi-CPU platform
 \, or servers in a geographically dispersed cluster. Insisting that such m
 ulti-node systems operate synchronously limits their scalability\, since t
 he performance is then dictated by the slowest node\, and the system becom
 es fragile to node failures. Hence\, there is a strong current interest in
  developing asynchronous algorithms for optimal decision-making. However\,
  the dynamics of asynchronous iterations are much richer than their synchr
 onous counterparts and quantifying the impact of asynchrony on the converg
 ence rate is mathematically challenging. \n\nIn this talk\, I will descri
 be some of our recent efforts in analyzing and designing asynchronous opti
 mization algorithms with strong convergence guarantees.\n \nI will begin 
 by introducing novel convergence results for asynchronous iterations that 
 appear in the analysis of parallel and distributed optimization algorithms
 . The results are simple to apply and give explicit estimates for how the 
 degree of asynchrony impacts the convergence rates of the iterates. Our re
 sults shorten\, streamline and strengthen existing convergence proofs for 
 several asynchronous  optimization methods and allow us to establish conv
 ergence guarantees for popular algorithms that were thus far lacking a com
 plete theoretical understanding. \n \nI will then proceed to discuss how
  one can design delay-adaptive optimization algorithms. In contrast to ear
 lier efforts that used fixed step-sizes based on the worst-case delay\, th
 ese algorithms measure the actual information delays in the system and ada
 pt the step-sizes on-line to maintain a fast and stable convergence. We wi
 ll also discuss alternative mechanisms that use fixed step-sizes but drop 
 information that is considered too old.\n \nFinally\, if time permits\, I
  will discuss classes of problems that admit delay-agnostic algorithms. Su
 ch algorithms can be tuned without knowledge of the level of asynchrony th
 at they will experience when they are deployed. Our analysis proves that t
 hese algorithms converge for all levels of asynchrony\, and quantifies how
  the convergence times increase as the information delays grow larger.\n 
 \nThis talk is based on joint work with Hamid Reza Feyzmahdavian\, Xuyang 
 Wu\, Sindri Magnusson and Changxin Liu. \n \nBio: Mikael Johansson  re
 ceived the M.Sc. and Ph.D. degrees in electrical engineering from Lund Uni
 versity\, Sweden\, in 1994 and 1999\, respectively. He held postdoctoral p
 ositions with Stanford University and  the University of California at Be
 rkeley before joining the KTH Royal Institute of Technology\, Stockholm\, 
 Sweden\, in 2002\, where he is currently a Full Professor. He has played a
  leading role in several national and international research projects on o
 ptimization\, control\, and communications. He has coauthored two books an
 d over 200 research papers\, several of which are highly cited and have re
 ceived recognition in terms of best paper awards. His research revolves ar
 ound large-scale and distributed optimization\, autonomous decision making
 \, control\, and machine learning. Dr. Johansson has served on the Editori
 al Boards of Automatica and the IEEE Transactions on Network Systems and o
 n the program committee for several top conferences in control\, communica
 tions\, and machine learning. \n\n\nwebsite: https://people.kth.se/~mika
 elj/\n\nNCCR Automation: https://nccr-automation.ch/ 
LOCATION:MA B1 11 https://plan.epfl.ch/?room==MA%20B1%2011 https://epfl.zo
 om.us/j/9981475369?omn=62709738811
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR