IMPERIAL COLLEGE STATISTICS SECTION SEMINARS
Friday 5th November
Discretisation for Inference on Bayesian Mixture Models
----Mark Brewer (Exeter University)
To be held at 2pm in Room 140 Huxley on the South Kensington campus.
Abstract:
The problem of inference on Bayesian mixture models is known to be
difficult; direct estimation of posterior quantities is hampered
by a combinatorial explosion which requires every possible partition
of the data set to be considered. Most current work focuses on MCMC
analysis which avoids this combinatorial pitfall; this talk discusses
an alternative strategy which discretises the prior distributions and
then uses an efficient summation to reduce the computing time from the
usual O(2^n) (for a 2-component mixture) to O(D^2 * n^2), where D is
the level of discretisation. In this way, direct estimation is made
feasible, and accuracy is constrained only by the discretisation of
the priors.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|