BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Statistical Laboratory Graduate Seminars
SUMMARY:Bayesians turn to Experts for Advice! - Steven de
Rooij (Cambridge)
DTSTART;TZID=Europe/London:20081119T170000
DTEND;TZID=Europe/London:20081119T180000
UID:TALK14680AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/14680
DESCRIPTION:*In Bayesian model selection and model averaging
\, inference is normally based on a posterior dist
ribution on the models\, usually\n interpreted as
a measure of how likely we consider each of the mo
dels to\n be "true"\, or at least in some sense cl
ose to true\, given on the\n observations.\n R
ather than with truth\, I will be concerned with t
he more practical\n goal of finding a "useful" mod
el\, in the sense that it predicts future\n outcom
es of the underlying process well. As it turns out
\, the most\n useful model may well vary depending
on the number of available\n observations! For in
stance\, given ten samples from some continuous\n
density\, a seven-bin histogram model is more usef
ul than a 1\,000-bin\n model\, even though the lat
ter is arguably closer to being "true".\n As i
t turns out\, methods for tracking transient perfo
rmance of\n prediction strategies have already bee
n developed in the learning theory\n literature un
der the heading "prediction with expert advice". I
will\n illustrate how these methods can improve m
odel selection performance\n using results from co
mputer simulations on density estimation problems.
*\n
LOCATION:CMS\, MR4
CONTACT:HoD Secretary\, DPMMS
END:VEVENT
END:VCALENDAR