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:Inference Group
SUMMARY:Coherent Inference on Optimal Play in Games - Phil
ipp Hennig (University of Cambridge)
DTSTART;TZID=Europe/London:20100510T110000
DTEND;TZID=Europe/London:20100510T120000
UID:TALK24617AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/24617
DESCRIPTION:The search for an optimal path through a game tree
is one of the oldest problems in computer science
. Over the past years\, Monte Carlo tree search ha
s emerged as a surprisingly effective approach to
this problem. I will present a probabilistic gener
ative model for game trees that explains why MC tr
ee search works at all. I will then move on to der
ive an approximate inference algorithm for this mo
del\, which can infer beliefs over the value of _a
ny_ node in the tree under _optimal_ play\, using
random roll-out data from other parts in the tree.
Somewhat surprisingly\, this inference algorithm
is of linear complexity\, even though the exact se
arch problem has exponential cost.\n\nThe work pre
sented in this talk has just been published as P.
Hennig\, D. Stern\, T. Graepel: "Coherent Inferenc
e on Optimal Play in Game Trees"\, J Machine Learn
ing Research\, W&CP 9 (2010)\, 326-333. \n\nSee\nh
ttp://jmlr.csail.mit.edu/proceedings/papers/v9/hen
nig10a/hennig10a.pdf
LOCATION:TCM Seminar Room\, Cavendish Laboratory\, Departme
nt of Physics
CONTACT:Emli-Mari Nel
END:VEVENT
END:VCALENDAR