BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1225
DTSTAMP:20230914T125955Z
SUMMARY:Learning Optimal Bids in Second Price Auctions with Temporal and Ov
 erlapping Targeting Constraints
DESCRIPTION:Speaker: Ravi R. Mazumdar (University of Waterloo)\n\nAbstract:
  \nAd placement in web-browsing and wireless mobiles is an increasingly im
 portant component of the advertisement market. The market size is over $ 1
 00 billion and counting. The mechanism is as follows: when a user opens a 
 webpage or mobile ap a message is sent to an exchange where multiple bidde
 rs have the possibility of placing an ad that would target the right user\
 , eg. age\, sex\, location\, etc. The ad that is displayed corresponds to 
 the bidder who bids the highest while the cost is calculated according to 
 a first or second price. Typically bidders are DSP (Demand Side Platforms)
  that aggregate bids on behalf of clients who wish to run a campaign for a
  given length of time with certain targeting criteria. The goal is to mini
 mize the total cost of obtaining the required number of impressions (ads t
 hat meet targeting criteria) over the duration of a contract. The real tim
 e constraint is that bidding must be done within 100ms.\n\nIn this talk I 
 will build upon the theory that we had earlier developed for computing the
  least cost bids in the second price context. This involves the notion of 
 an information state for the problem. There is a very rich primal-dual the
 ory that emerges\, one in the so called impressions space and the other in
  the contracts space. Computationally and structurally the primal and dual
  views of the optimization have different properties that can be exploited
  to come up with fast algorithms.\n\nThe optimal solutions depend on solvi
 ng a constrained convex optimization problem when the information state is
  known. However this is not readily available and thus there is the proble
 m of learning the information state. We show that in the second price case
 \, stochastic approximation (SA) algorithms that operate on censored data 
 (prices are only known by a bidder when the bidder wins) can be devised th
 at solve the constrained optimization problem without learning the informa
 tion state explicitly and we prove their convergence. Finally I will prese
 nt the dynamic behaviour through simulations.\nJoint work with Ryan Kinnea
 r (Waterloo) and Peter Marbach (Toronto). We thank Addictive Mobility Inc.
 \, a Pelmorex company for having proposed the problem and to Addictive Mob
 ility\, Ontario OCE VIP II\, and NSERC funding the work.\n\nBio: The speak
 er was educated at the Indian Institute of Technology\, Bombay (B.Tech\, 1
 977)\, Imperial College\, London (MSc\, DIC\, 1978) and obtained his PhD i
 n Control Theory under A. V. Balakrishnan at UCLA in 1983. He is currently
  a University Research Chair Professor in the Dept. of ECE at the Universi
 ty of Waterloo\, Ont.\, Canada where he has been since September 2004. Pri
 or to this he was Professor of ECE at Purdue University\, West Lafayette\,
  USA. Since 2012 he is a D.J. Gandhi Distinguished Visiting Professor at t
 he Indian Institute of Technology\, Bombay\, India and since May 2019 an A
 djunct Professor at the Tata Institute of Fundamental Research (TIFR)\, Mu
 mbai. He is a Fellow of the IEEE and the Royal Statistical Society. He is 
 a recipient of the INFOCOM 2006 Best Paper Award\, the ITC-27 2015 Best Pa
 per Award\, the Performance 2015 Best Paper Award and was runner-up for th
 e Best Paper Award at INFOCOM 1998. His research interests are in stochast
 ic modelling and analysis applied to complex networks and statistical infe
 rence.\nYouTube Link: https://youtu.be/dA44DA8gQDM\n
URL:https://www.tcs.tifr.res.in/web/events/1225
DTSTART;TZID=Asia/Kolkata:20220725T160000
DTEND;TZID=Asia/Kolkata:20220725T170000
LOCATION:A-201 and Zoom
END:VEVENT
END:VCALENDAR