Presented by:
Jonathan Goodman
Date:
Wednesday 20th November 2019 - 15:05 to 15:35
Venue:
INI Seminar Room 2
Event:
Abstract:
ABSTRACT: MCMC sampling
can use ideas from the optimization community. Optimization via Newton’s
method can fail without line search, even for smooth strictly convex problems.
Affine invariant Newton based MCMC sampling uses a Gaussian proposal
based on a quadratic model of the potential using the local gradient and
Hessian. This can fail (conjecture: give a transient Markov chain) even
for smooth strictly convex potentials. We describe a criterion that
allows a sequence of proposal distributions from X_n with decreasing “step
sizes” until (with probability 1) a proposal is accepted. “Very detailed
balance” allows the whole process to preserve the target distribution.
The method works in experiments but the theory is missing.