BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Timo Berthold (FICO)
DTSTART:20200630T180000Z
DTEND:20200630T183000Z
DTSTAMP:20260423T021932Z
UID:DOTs/17
DESCRIPTION:Title: <a href="https://researchseminars.org/talk/DOTs/17/">Le
 arning to Scale</a>\nby Timo Berthold (FICO) as part of Discrete Optimizat
 ion Talks\n\n\nAbstract\nScaling is a widely used preconditioning techniqu
 e\, used to reduce error propagation and thereby improve the numerical beh
 avior of an algorithm.\nFor numerically challenging mixed-integer programs
  (MIPs)\, as they appear in many practical applications\, having an effici
 ent scaling method in place often makes the difference whether the MIP's L
 P relaxations can be solved gracefully or not.\nThere are two scaling meth
 ods which are commonly used: Standard scaling and Curtis-Reid scaling.\nTh
 e latter often\, but not always\, leads to a more robust solution process\
 , but also to longer solution times.\nWe introduce a method to automatical
 ly choose between the two scaling variants by predicting which one will le
 ad to fewer numerical issues.\nIt turns out that this not only reduces var
 ious types of numerical errors\, but is also performance-neutral for MIPs 
 and improves performance on LPs.\n
LOCATION:https://researchseminars.org/talk/DOTs/17/
END:VEVENT
END:VCALENDAR