BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.bham.ac.uk//v3//EN
BEGIN:VEVENT
CATEGORIES:Artificial Intelligence and Natural Computation se
minars
SUMMARY:Bayesian Machine Learning for Controlling Autonomo
us Systems - Dr. Marc Deisenroth\, Imperial Colleg
e London
DTSTART:20130911T100000Z
DTEND:20130911T110000Z
UID:TALK1163AT
URL:/talk/index/1163
DESCRIPTION:Autonomous learning has been a promising direction
in control and\nrobotics for more than a decade s
ince learning models and controllers\nfrom data al
lows us to reduce the amount of engineering knowle
dge that\nis otherwise required.\nDue to their fle
xibility\, autonomous reinforcement\nlearning (RL)
approaches typically require many interactions wi
th the\nsystem to learn controllers. However\, in
real systems\, such as robots\,\nmany interactions
can be impractical and time consuming. To address
this\nproblem\, current learning approaches typic
ally require task-specific\nknowledge in form of e
xpert demonstrations\, pre-shaped policies\, or\ns
pecific knowledge about the underlying dynamics.\n
\nIn the first part of the talk\, we follow a diff
erent approach and speed\nup learning by efficient
ly extracting information from sparse data. In\npa
rticular\, we learn a probabilistic\, non-parametr
ic Gaussian process\ndynamics model. By explicitly
incorporating model uncertainty into\nlong-term p
lanning and controller learning our approach reduc
es the\neffects of model errors\, a key problem in
model-based learning. Compared\nto state-of-the a
rt RL our model-based policy search method achieve
s an\nunprecedented speed of learning. We demonstr
ate its applicability to\nautonomous learning in r
eal robot and control tasks.\n\nIn the second part
of my talk\, we will discuss two alternative meth
ods\nfor learning controllers: a) imitation learni
ng and b) Bayesian\noptimization. In imitation lea
rning\, it is no longer necessary to\nspecify a ta
sk-dependent cost function. Instead\, a teacher de
monstrates\na task\, which the robot should imitat
e. I will show that probabilistic\nmodels are very
useful for rapid imitation learning. Bayesian\nop
timization is typically used to optimizes expensiv
e-to-evaluate\nfunctions. We successfully applied
Bayesian optimization to learning\ncontroller para
meters for a bipedal robot\, where modeling the dy
namics\nis very difficult due to ground contacts.
Using Bayesian optimization\,\nwe sidestep this mo
deling issue and directly optimize the controller\
nparameters without the need of modeling the robot
's dynamics.\n\n==== Tea and cookies just after th
e seminar in the coffee room ====\n
LOCATION:Mech Eng G26
CONTACT:Leandro Minku
END:VEVENT
END:VCALENDAR