BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Splitting algorithms: efficient lightweight solvers for real-time 
 embedded nonlinear MPC
DTSTART:20200313T110000
DTEND:20200313T120000
DTSTAMP:20260407T014828Z
UID:75521b348c0e354e3dd63749a0ab8925a5b1e7f842b2c06f4a57e3a3
CATEGORIES:Conferences - Seminars
DESCRIPTION:Andreas Themelis - KU Leuven\, Belgium\nModel predictive contr
 ol (MPC) has become a popular strategy to implement feedback control loops
  for a variety of systems. Since most systems are nonlinearby nature\, n
 onlinear MPC (NMPC) offers a more accurate modeling than linear MPC\, b
 ut it leads to nonconvex and much more complicated problems. At every samp
 ling step\, NMPC requires the solution of a nonlinear program (NLP) tha
 t has to be computed within sampling time\, thus imposing an imperative de
 mand for algorithmic speed and efficiency.\n\nA usual approach to this typ
 e of problems is sequential quadratic programming (SQP)\, which requires t
 he solution of a quadratic program at every iteration and\, consequently\,
  inner iterative procedures. As a result\, each outer iteration may become
  computationally very expensive. This constitute a severe drawback for all
  those applications in which the solver is embedded on a simple platform w
 ith low memory and limited computational power. On the contrary\, "splitti
 ng algorithms" such as the proximal (projected) gradient typically involve
  elementary operations and negligible memory allocation\, but are extremel
 y sensitive to problem conditioning and may require prohibitively many ite
 rations before converging to a satisfactory solution.\n\nIn this talk we c
 ombine basic favorable properties of the forward-backward envelope (FBE) t
 o design a globally convergent\, fast\, and lightweight algorithm perfectl
 y suited for embedded applications. The algorithm\, named PANOC (Proximal 
 Averaged Newton method for Optimal Control)\, is of linesearch type and co
 mbines the global convergence of the projected gradient method with the fa
 st local behavior of L-BFGS-type directions\, yet maintaining the operatio
 nal simplicity of the former method. Unlike classical linesearch methods\,
  it does not require differentiability assumptions nor is it limited to "d
 escent directions"\, and thus differs substantially from early FBE-based m
 inimization algorithms. The proposed methodology is validated on a lab-sca
 le vehicle that has to autonomously move in an obstructed environment to a
  reference position. Numerical results show that PANOC can be up to two or
 ders of magnitude faster than interior point and SQP-based solvers\, as we
 ll as previous FBE-based minimization algorithms operating a classical lin
 esearch.\n\nBio: \nAndreas Themelis is a postdoctoral researcher at the De
 partment of Electrical Engineering (ESAT) of KU Leuven\, Belgium\, since J
 anuary 2019. AGer obtaining MSc and BSc degrees in MathemaJcs from the Uni
 versity of Florence (Italy)\, he received a joint PhD degree in Electrical
  Engineering from KU Leuven (Belgium) and in Control\, Decision and System
  Sciences from IMT Lucca (Italy) in December 2018. His research focuses on
  the convergence analysis and Newton-type acceleraJon of spliUng algorithm
 s for nonsmooth and nonconvex problems\, aiming at developing high-level c
 ompeJJve algorithms with low computaJonal and memory requirements that can
  run on simple plaXorms and are thus well suited for embedded applicaJons.
 \n 
LOCATION:ME C2 405 https://plan.epfl.ch/?room==ME%20C2%20405
STATUS:CANCELLED
END:VEVENT
END:VCALENDAR