BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VTIMEZONE
TZID:Europe/Vienna
BEGIN:DAYLIGHT
DTSTART:20190331T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20181028T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20191018T063657Z
UID:5c594f34b8dd8802140041@ist.ac.at
DTSTART:20190207T103000
DTEND:20190207T113000
DESCRIPTION:Speaker: Emir Demirovic\nhosted by Vladimir Kolmogorov\nAbstrac
t: In the first part of the talk\, we introduce the Predict+Optimise probl
em setting and discuss developing machine learning algorithms that are spe
cifically designed to be used with combinatorial problems. These algorithm
s are important when machine learning and optimisation must interact. For
instance\, when optimisation is to be performed on data that is not entire
ly correct but is rather estimated with machine learning. The main challen
ge is that conventional machine learning metrics\, such as mean-square err
or\, are not necessarily indicative of the outcome of optimisation procedu
re. Ideally\, the latter would be used as a criterion of success. However\
, not only does this involve solving an (NP-hard) optimisation problem\, b
ut in addition\, this gives riseto a nonlinear\, noncontinuous\, and nondi
fferentiable function. Hence\, techniques widely used in machine learning\
, e.g. gradient descent\, cannot be applied. The standard approach is to v
iew optimisation and machine learning as two separate black boxes\, but po
tentially better results can be obtained if the process is merged into a s
ingle pipeline. Although both combinatorial optimisation and machine learn
ing have been thoroughly studied\, their interplay has yet to be understoo
d. In this talk\, we present ourformalisation of the Predict+Optimise prob
lem and its challenges\, discuss various approaches\, and provide an exper
imental study.The second part of the talk is devoted to the maximum Boolea
n satisfiability problem (MaxSAT)\, where the aim is to compute an assignm
ent of variables that satisfy as many clauses as possible for a propositio
nal logic formula. Many real-life problems can be formulated in propositio
nal logic and thus developing algorithms for this problem is of high impor
tance. Given that practical applications can lead to large formulas that n
eed to be solved with tight time budgets\, we focus our attention on metho
ds that provide 'good' solutions 'quickly\; We developed techniques for Ma
xSAT solving based on the exhaustive linear search algorithm and utilise t
echniques inspired by local search. The resulting algorithm proved to be h
ighly effective\, taking first place in the 300 seconds incomplete track o
f the MaxSAT Evaluation 2018. Moreover\, we briefly discuss further improv
ements based on the integration of the algoithm with a core-guided approac
h\, essentially merging lower and upper bounding techniques.Dr Emir Demiro
vi? is an associate lecturer and postdoctoral researcher at the University
of Melbourne in Australia\, working closely with Professors Peter J. Stuc
key\, James Bailey\, Rao Kotagiri\, Christopher Leckie\, and Dr Jeffrey Ch
an. His main expertise lies in solving combinatorial optimisation problems
through the use of constraint/integer programming\, metaheuristics\, and
the combination thereof\, with special attention to timetabling and schedu
ling algorithms. After earning his Ph.D. in 2017 at TU Wien under Priv.-Do
z. Dr. Nysret Musliu supervision\, Dr Demirovi? worked as an optimisation
expert in the industry prior to coming to Australia. Throughout his career
\, he maintained a strong collaboration with the National Institute of Inf
ormatics and the National Institute of Advanced Industrial Science and Tec
hnology in Tokyo\, where he was invited several times. More information ca
n be found on his personal website:www.emirdemirovic.com.
LOCATION:Meeting room 3rd floor / Central Bldg. (I01.3OG.Meeting Room)\, IS
T Austria
ORGANIZER:abonvent@ist.ac.at
SUMMARY:Combinatorial Machine Learning and Incomplete MaxSAT Solving
URL:https://talks-calendar.app.ist.ac.at/events/1796
END:VEVENT
END:VCALENDAR