BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:America/New_York
BEGIN:DAYLIGHT
DTSTART:20250309T030000
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3
TZNAME:EDT
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251102T010000
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11
TZNAME:EST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T182351Z
UID:721EE321-E3EA-42FC-9249-5FA369694167
DTSTART;TZID=America/New_York:20251024T130000
DTEND;TZID=America/New_York:20251024T150000
DESCRIPTION:Abstract: In this talk\, we draw an explicit analogy across fou
 r problem classes in optimization and control with a unified solution char
 acterization. This viewpoint allows for a systematic transformation of alg
 orithms from one domain to the other. With this in mind\, we exploit two l
 inear structural constraints specific to control problems with finite stat
 e-action pairs to approximate the Hessian in a second-order-type algorithm
  from optimization. This leads to novel first-order control algorithms wit
 h the same computational complexity as (model-based) value iteration and (
 model-free) Q-learning\, while they exhibit an empirical convergence behav
 ior similar to (model-based) policy iteration and (model-free) Zap Q-learn
 ing with very low sensitivity to the discount factor. If time permits\, we
  also discuss how an interesting analogy between the convex conjugate oper
 ator and the Fourier transform can reduce the typical time complexity of t
 he dynamic programming operation from O(XU) to O(X + U) where X and U deno
 te the size of the discrete state and input spaces\, respectively.\n\nSpea
 ker(s): Peyman Mohajerin Esfahani\, \n\nRoom: SF B650\, Bldg: MIE\, Univer
 sity of Toronto\, 172 St. George St.\, Toronto\, Ontario\, Canada\, M5R 0A
 3
LOCATION:Room: SF B650\, Bldg: MIE\, University of Toronto\, 172 St. George
  St.\, Toronto\, Ontario\, Canada\, M5R 0A3
ORGANIZER:mehrdad.tirandazian@ieee.org
SEQUENCE:19
SUMMARY:From Optimization to Control: An Algorithmic Perspective
URL;VALUE=URI:https://events.vtools.ieee.org/m/509044
X-ALT-DESC:Description: <br /><p><span style="font-size: 12pt\;"><strong st
 yle="color: rgb(34\, 34\, 34)\; font-family: Arial\, Helvetica\, sans-seri
 f\; font-style: normal\; font-variant-ligatures: normal\; font-variant-cap
 s: normal\; letter-spacing: normal\; text-align: start\; text-indent: 0px\
 ; text-transform: none\; word-spacing: 0px\; -webkit-text-stroke-width: 0p
 x\; white-space: normal\; background-color: rgb(255\, 255\, 255)\; text-de
 coration-thickness: initial\; text-decoration-style: initial\; text-decora
 tion-color: initial\;">Abstract:&nbsp\;</strong><span style="color: rgb(34
 \, 34\, 34)\; font-family: Arial\, Helvetica\, sans-serif\; font-style: no
 rmal\; font-variant-ligatures: normal\; font-variant-caps: normal\; font-w
 eight: 400\; letter-spacing: normal\; text-align: start\; text-indent: 0px
 \; text-transform: none\; word-spacing: 0px\; -webkit-text-stroke-width: 0
 px\; white-space: normal\; background-color: rgb(255\, 255\, 255)\; text-d
 ecoration-thickness: initial\; text-decoration-style: initial\; text-decor
 ation-color: initial\; display: inline !important\; float: none\;">In this
  talk\, we draw an explicit analogy across four problem classes in optimiz
 ation and control with a unified solution characterization. This viewpoint
  allows for a systematic transformation of algorithms from one domain to t
 he other. With this in mind\, we exploit two linear structural constraints
  specific to control problems with finite state-action pairs to approximat
 e the Hessian in a second-order-type algorithm from optimization. This lea
 ds to novel first-order control algorithms with the same computational com
 plexity as (model-based) value iteration and (model-free) Q-learning\, whi
 le they exhibit an empirical convergence behavior similar to (model-based)
  policy iteration and (model-free) Zap Q-learning with very low sensitivit
 y to the discount factor. If time permits\, we also discuss how an interes
 ting analogy between the convex conjugate operator and the Fourier transfo
 rm can reduce the typical time complexity of the dynamic programming opera
 tion from O(XU) to O(X + U) where X and U denote the size of the discrete 
 state and input spaces\, respectively.</span></span></p>
END:VEVENT
END:VCALENDAR